<?xml version="1.0" encoding="utf-8"?>
<feed xmlns="http://www.w3.org/2005/Atom" xml:lang="en">
<title>albadr.ai Blog</title>
<subtitle>The blog of albadr.ai: what BADR and Translator do, how they are put together, and what we measured.</subtitle>
<link href="https://blog.albadr.ai/feed.xml" rel="self"/>
<link href="https://blog.albadr.ai/"/>
<id>https://blog.albadr.ai/</id>
<updated>2026-11-17T08:00:00Z</updated>
<entry>
<title>Half a second behind the voice</title>
<link href="https://blog.albadr.ai/half-a-second-behind-the-voice/"/>
<id>https://blog.albadr.ai/half-a-second-behind-the-voice/</id>
<published>2026-11-17T08:00:00Z</published>
<updated>2026-11-17T08:00:00Z</updated>
<author><name>albadr.ai</name></author>
<summary>Translator is our MVP demo of live AI speech translation between five languages. On a desktop processor from 2012, with no graphics processor, an English word is on screen 573 milliseconds after it is spoken, on average, and about one recognized word in seven is wrong.</summary>
<content type="html">&lt;p&gt;In 2025 IWSLT, the research conference where live speech translation systems are evaluated, timed a professional interpreter with the stopwatch it uses for machines: one person, working alone, from the audio only. Interpreting Czech into English, the professional came out &lt;a href=&quot;https://aclanthology.org/2025.iwslt-1.44.pdf&quot;&gt;4.16 seconds&lt;/a&gt; behind the speaker by that metric, against 1.51 seconds for a research system on the same material. The human judges still rated the professional higher.&lt;/p&gt;
&lt;p&gt;Translator, our demo, puts a spoken English word on the screen 573 milliseconds after it is spoken, and its French translation covers that word after 607 milliseconds. Those are means over 99 recorded sentences, on a desktop processor from 2012 (an Intel Core i7-3770, four cores, 8 GB of memory) with no graphics processor in use.&lt;/p&gt;
&lt;p&gt;An interpreter produces speech and copes with real conversation. Our figures are for text on a screen, measured on recorded sentences that were read aloud, and about one recognized word in seven is wrong. Translator is an MVP demo, public since 15 November 2026 at &lt;a href=&quot;https://translate.albadr.ai&quot;&gt;translate.albadr.ai&lt;/a&gt;, with quotas.&lt;/p&gt;
&lt;h2 id=&quot;what-translator-is&quot;&gt;What Translator is&lt;/h2&gt;
&lt;p&gt;Translator is a web page. A person picks the language they speak and the one they want to read, presses Start and talks, and their words and the translation appear as text while they are still talking. It works in any of the 20 directions between English, French, Russian, Arabic and Chinese, five of the &lt;a href=&quot;https://www.un.org/en/our-work/official-languages&quot;&gt;six official languages of the United Nations&lt;/a&gt; (the sixth is Spanish). The output is text on a screen, not a synthetic voice.&lt;/p&gt;
&lt;p&gt;All of the work is done by AI: neural speech recognition and neural translation, several models at once for a language pair, in real time. None of them needs a graphics processor. The same models run on an ordinary processor in a small server, and inside a browser tab on the visitor’s own machine. Under its name the page says “Fully AI solution from albadr.ai”, and one of the cards on its first screen spells out what that means: “No person listens, types or translates. Like any AI, it can get a word wrong.” We can put a number on that last sentence. About one recognized word in seven is wrong on read sentences from many speakers, more with strong accents, noise or two people talking at once.&lt;/p&gt;
&lt;p&gt;As a demo it is rationed. A network address gets four hours a day with the engines on the visitor’s own device and twenty minutes a day through our server, and a session ends after 90 seconds without a voice. There are no accounts, no payment, no price and no interface for a client’s systems yet.&lt;/p&gt;
&lt;h2 id=&quot;where-an-interpreter-is-missing&quot;&gt;Where an interpreter is missing&lt;/h2&gt;
&lt;p&gt;A House of Lords committee reported &lt;a href=&quot;https://publications.parliament.uk/pa/ld5901/ldselect/pubserv/87/87.pdf&quot;&gt;8,958 unfulfilled requests for interpreters&lt;/a&gt; in the courts of England and Wales between January and September 2024. Where an interpreter can be found, the day has a price. United States federal courts pay a federally certified contract interpreter &lt;a href=&quot;https://www.uscourts.gov/court-programs/federal-court-interpreters&quot;&gt;$617 for a full day&lt;/a&gt;, and for simultaneous work interpreters &lt;a href=&quot;https://aiic.ch/faq/&quot;&gt;usually work in teams of two and relieve each other every 20 to 30 minutes&lt;/a&gt;, says the Swiss region of AIIC, their international association.&lt;/p&gt;
&lt;p&gt;We quote the courts because they keep count. They are the last place for a demo with that error rate. The need we have in mind is lower in stakes and, we think, far more common. In the 2021 census of England and Wales, &lt;a href=&quot;https://www.ons.gov.uk/peoplepopulationandcommunity/culturalidentity/language/bulletins/languageenglandandwales/census2021&quot;&gt;880,000 residents could not speak English well and 161,000 could not speak it at all&lt;/a&gt;, whatever their language. Some of them explain what they came for at a front desk, or sit in a clinic’s reception on a day when no interpreter was booked. It is the same at a lecture, a town-hall meeting, a guided visit, a sermon or on a support line: one person speaks, others must follow, and nobody was going to pay for a team of two interpreters. These are our proposals. Nothing is deployed in any of them.&lt;/p&gt;
&lt;p&gt;On how much delay people tolerate we can only quote others. People in conversation answer fast: in recordings of informal talk in ten languages, the answer to a yes-or-no question began &lt;a href=&quot;https://pmc.ncbi.nlm.nih.gov/articles/PMC2705608/&quot;&gt;208 milliseconds on average&lt;/a&gt; after the question ended, and &lt;a href=&quot;https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2015.00731/full&quot;&gt;a pause of 600 milliseconds or longer&lt;/a&gt; before a reply makes listeners expect an unwelcome answer, according to a review of research on turn-taking. For a translated voice, Google’s Meet team has said that earlier chains of transcription, translation and speech synthesis were often 10 to 20 seconds late, and that &lt;a href=&quot;https://blog.google/products-and-platforms/products/workspace/google-meet-langauge-translation-ai/&quot;&gt;“two to three seconds was sort of a sweet spot”&lt;/a&gt;. That is a product manager quoted on the company’s blog, about a voice, with no method given. We have not studied readers of our own page, and we claim nothing about comprehension at half a second.&lt;/p&gt;
&lt;h2 id=&quot;one-design-that-runs-in-two-places&quot;&gt;One design that runs in two places&lt;/h2&gt;
&lt;p&gt;The page captures the microphone and draws the text. It and the engine files are plain files, served without any program running. A small program at the network edge answers the page’s questions and admits sessions, and a service keeps the usage budget. The engines, the neural models that recognize speech and translate it, are the only heavy part, and they can run in two places.&lt;/p&gt;
&lt;p&gt;One is the visitor’s own device. The engines are downloaded once (218 MB for English to French, 213 to 451 MB for the other pairs) and run in the browser. In this mode the audio does not leave the browser at all, which matters because a voice says more than its words: the French data protection authority CNIL lists &lt;a href=&quot;https://www.cnil.fr/sites/cnil/files/atoms/files/cnil_white-paper-on_the_record.pdf&quot;&gt;age, gender, health or emotional state and identity&lt;/a&gt; among what it reveals. What does travel to our service is text and numbers, the recognized words and their translation among them. The service handles them in memory and stores none of it. What it keeps is a ledger of the time used and notes on its own faults, and its logging is switched off. This mode needs the download and is only as fast as the machine. Each device times itself before the large download and is told if it would not keep up, and an older tablet browser that cannot run the engines is sent to the server.&lt;/p&gt;
&lt;p&gt;The other is a small container with 2 processors and 3.5 GiB of memory. Audio is streamed to it and exists only in memory. No audio is written to disk or logged. The container sleeps when nobody uses it, and the first visitor after a sleep pays for that: on the live site, a start from sleep opens its connection in about a second and has its models loaded after seven. A visitor who was promised that the microphone stays on the device is asked before any audio goes to the server.&lt;/p&gt;
&lt;p&gt;Doing without a graphics processor is what lets the server fit in a container of that size and the device side in a browser tab. The cost is that models small enough for a processor hear like small models.&lt;/p&gt;
&lt;p&gt;The page also reports on itself. Every message of live text carries the moment its newest word was spoken, on the page’s own clock, so the delay it shows is the delay that user experienced, transport included. Text that may still change is drawn lighter with a dotted underline. The choice was between showing words early and never taking a word back. We show early and mark every change.&lt;/p&gt;
&lt;p&gt;The whole public demo is designed to stay inside one hosting plan of 5 USD a month, under a budget the service enforces. A visitor is counted under a name that is a salted digest of the network address. When a share is used up the page says so, and says when it is worth coming back. Turning a visitor away until a stated time seemed better than an unplanned bill.&lt;/p&gt;
&lt;h2 id=&quot;what-we-measured&quot;&gt;What we measured&lt;/h2&gt;
&lt;p&gt;The figures in the table come from the test part of FLEURS, a public set of read sentences that no setting was chosen on. The English column is a single run, of 11 November. The French column is the last of six runs made from 11 to 14 November while the French models were still being changed, and its word errors read between 13.3% and 15.4% along the way, so it is not as clean a held-out figure as the English one. All were played into the real system at the speed of speech, on the 2012 desktop processor, with the engines on the same machine and no network in between.&lt;/p&gt;
&lt;div class=&quot;table&quot;&gt;&lt;table&gt;
&lt;thead&gt;&lt;tr&gt;&lt;th&gt;&lt;/th&gt;&lt;th class=&quot;to-end&quot;&gt;English to French&lt;/th&gt;&lt;th class=&quot;to-end&quot;&gt;French to English&lt;/th&gt;&lt;/tr&gt;&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;&lt;td&gt;Recordings&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;99&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;100&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;A spoken word is on screen after, mean (95th percentile)&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;573 ms (715)&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;532 ms (683)&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;That word is no longer marked provisional, mean (95th percentile)&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;836 ms (1126)&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;791 ms (1085)&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;Its translation covers it after, mean (95th percentile)&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;607 ms (755)&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;574 ms (740)&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;Whole translation on screen after the voice stops, median (95th percentile)&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;470 ms (684)&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;410 ms (661)&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;Word errors in the text as it finally stands&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;14.10%&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;13.63%&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;Translation as it finally stands, BLEU and chrF&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;36.18 and 63.39&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;31.86 and 61.15&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;Settled translation words replaced when a sentence is declared final&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;0.88%&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;0.33%&lt;/td&gt;&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;
&lt;p&gt;Read the second row with the third. “On screen” includes text still marked provisional, and the third row says when the mark is gone. Computing is the small part: one step of recognition and one translation together take under 0.1 s on average on that processor. Most of the half second is the recognizer needing to hear what follows a word before it can write it.&lt;/p&gt;
&lt;p&gt;The error and translation figures are for the text as it finally stands, after any revision. In French a revision arrives 888 ms on average after the voice stops (95th percentile 1,571 ms), and until then about one French word in five is wrong. On 115 Arabic recordings that no setting was chosen on, the text as it finally stands on the server has 6.9% word errors. The last row counts only replacements when a sentence is declared final. While a sentence is still open, 16 to 20% of the settled translation words change, and the page marks each change.&lt;/p&gt;
&lt;p&gt;Seven more directions were measured only on development recordings, the ones our settings were tuned on, 19 to 60 per direction. In those runs a word is on screen after 539 to 612 ms. With a simulated network round trip of 60 ms, the English to French figure was 669 ms through the server (42 development recordings).&lt;/p&gt;
&lt;p&gt;Device mode is slower on the same machine: with the engines’ browser builds standing in for a device and the same simulated network, 644 to 840 ms across the 20 directions (17 to 20 development recordings each). For English, accuracy held in a browser. In a proof on one PC and one browser, the English recognizer made 11.64% word errors in a tab against 11.48% natively on 60 held-out recordings (46 of the 60 transcripts were identical), and the translation model, given the same 60 sentences, returned the server’s translations. Arabic is different: on the device stand-in (17 to 20 development recordings a direction) it makes 11.6 to 13.1% word errors against 5.9 to 7.2% on the server, until the visitor switches on an option that fetches 174 MB more.&lt;/p&gt;
&lt;p&gt;We did not measure spontaneous speech, noise, strong accents, two people talking at once, phones, real networks, or the speed of the hosted container. Russian, Arabic and Chinese have no held-out delay figures. No person has rated the translations.&lt;/p&gt;
&lt;h2 id=&quot;what-the-industry-ships-and-charges&quot;&gt;What the industry ships and charges&lt;/h2&gt;
&lt;p&gt;Large vendors ship live speech translation today, each in more languages than our five. Google Meet &lt;a href=&quot;https://workspaceupdates.googleblog.com/2026/10/real-time-speech-translation-in-google-Meet-expands-to-63-languages-with-automatic-language-detection.html&quot;&gt;translates speech between any combination of 63 languages&lt;/a&gt; on eligible plans, and its help page says translations &lt;a href=&quot;https://support.google.com/meet/answer/16221730?hl=en&amp;amp;co=GENIE.Platform%3DDesktop&quot;&gt;“will be a few seconds delayed”&lt;/a&gt;. Apple’s Live Translation with AirPods does &lt;a href=&quot;https://support.apple.com/en-us/123185&quot;&gt;all processing on the iPhone&lt;/a&gt;, needs an iPhone 15 Pro or later, and has &lt;a href=&quot;https://www.apple.com/ios/feature-availability/&quot;&gt;neither Russian nor Arabic on its list&lt;/a&gt;. Microsoft’s Interpreter in Teams &lt;a href=&quot;https://support.microsoft.com/en-us/teams/copilot/interpreter-in-microsoft-teams-meetings-and-calls&quot;&gt;sends meeting audio to the cloud&lt;/a&gt;. Neither Apple’s nor Microsoft’s page gives a delay in seconds.&lt;/p&gt;
&lt;p&gt;Research publishes more. The IWSLT evaluation has &lt;a href=&quot;https://iwslt.org/2026/simultaneous&quot;&gt;a low-latency regime of 0 to 2 seconds and a high one of 2 to 4&lt;/a&gt;, by a metric that leaves computing time out, and expects submitted systems to run on a single NVIDIA H100 graphics processor with 80 GB. Its 2026 findings put computing at &lt;a href=&quot;https://aclanthology.org/2026.iwslt-1.39.pdf&quot;&gt;about 0.3 seconds more&lt;/a&gt; where that could be measured. None of this is like for like. Those systems are tested on other material and by other metrics. We draw no league table. All we claim is that a delay of this size did not require a graphics processor.&lt;/p&gt;
&lt;p&gt;On 17 November 2026 Microsoft’s Azure listed real-time speech translation at &lt;a href=&quot;https://azure.microsoft.com/en-us/pricing/details/cognitive-services/speech-services/&quot;&gt;$2.50 per audio hour&lt;/a&gt;. That is a price to a customer, and we have none to set beside it. What we have is a hosting cost. By our own simulation, not from traffic, the 5 USD plan carries about 470 hours of sessions a month on visitors’ devices and 10.3 hours on the server. Nobody had visited when those figures were computed, and we have seen no invoice. The device figure is large because the computing happens on the visitor’s processor: a device session costs the host only a few short messages a second while someone speaks, and none in silence.&lt;/p&gt;
&lt;h2 id=&quot;how-a-client-would-integrate-it-and-where-it-has-run&quot;&gt;How a client would integrate it, and where it has run&lt;/h2&gt;
&lt;p&gt;Today the demo is closed on purpose. The page cannot be embedded in another site, and calls from other sites are refused. On request we can offer a client the connection underneath. It is one streaming connection with a short vocabulary: audio frames go up, and events come down (ready, live text, final text, revised text). A client’s display has to do with those events what our page does: show live text as text that may still change, then the final text, then a revision if one arrives. The server side runs as a single process or as a container image, on a processor and nothing else. Device mode is there for clients who want the audio never to leave the user’s machine. Invitation links carry their own minutes, outside the public quotas.&lt;/p&gt;
&lt;p&gt;The rest would be built with the client, and we list it as work because none of it exists: keys and per-client metering, permission to embed, connectors to the client’s meeting, telephony or event systems, more languages, and output beyond text on a page.&lt;/p&gt;
&lt;p&gt;On 11 November 2026 we took the first measurements, with the whole product as one process on a desktop PC. A day later the same engines were proven in a browser tab, and the edge version ran locally. By 13 November all 20 directions ran with the engines’ browser builds on that PC, six of them in a real browser tab. The public launch followed on 15 November, on an edge platform plus a container, with all five languages on the server the same evening. So between 11 and 15 November it ran as one process on a PC, in a browser tab, in an edge runtime and in a Linux container. A new server has to offer a processor and room for one process or one container. We have no recorded run on another cloud, and no figures from phones. Changes are checked against seven reference traces (5,496 events) that must come out the same byte for byte, beside more than 270 automated tests.&lt;/p&gt;
&lt;h2 id=&quot;what-the-demo-cannot-do-yet&quot;&gt;What the demo cannot do yet&lt;/h2&gt;
&lt;p&gt;Translator is an MVP demo. About one recognized word in seven is wrong on read sentences, and we have no figures for spontaneous speech, noise, accents or phones, nor for delay over a real network. It was tested with one speaker and one microphone, and overlapping speakers are not handled. There are five languages, quotas and no price. Device mode begins with a download of 213 to 451 MB, and an older tablet browser falls back to the server. The interface exists in all five languages, in translations that native speakers have not yet read. The capacity figures come from a simulation.&lt;/p&gt;
&lt;p&gt;The demo is at &lt;a href=&quot;https://translate.albadr.ai&quot;&gt;translate.albadr.ai&lt;/a&gt;. Speak a sentence, then open the settings panel and read the delay the page measured for you. If you have a room where one person speaks and others must follow, or a system this should connect to, write to contact@albadr.ai.&lt;/p&gt;</content>
</entry>
<entry>
<title>The tab nobody opened</title>
<link href="https://blog.albadr.ai/the-tab-nobody-opened/"/>
<id>https://blog.albadr.ai/the-tab-nobody-opened/</id>
<published>2026-11-17T08:00:00Z</published>
<updated>2026-11-17T08:00:00Z</updated>
<author><name>albadr.ai</name></author>
<summary>A Word, Excel, PowerPoint or PDF file holds more than the page it shows. BADR is our MVP demo for the minute before a file is sent, and this is how it is put together, what we measured, and what it cannot do yet.</summary>
<content type="html">&lt;p&gt;On 3 February 2025 a manager at the Canada Border Services Agency sent a shift-scheduling spreadsheet to 70 employees. The file also contained &lt;a href=&quot;https://www.priv.gc.ca/en/opc-actions-and-decisions/investigations/investigations-into-federal-institutions/2025-26/pa_20260226_cbsa_01/#:~:text=secondary%20tab&quot;&gt;a secondary tab with personal information on more than 18,000 other employees&lt;/a&gt;. A workbook went out by email, and it was larger than the sheet its sender was looking at. The agency later found four other similar spreadsheet incidents in 2024 and 2025.&lt;/p&gt;
&lt;p&gt;It had happened before. On 8 August 2023 the Police Service of Northern Ireland answered a freedom-of-information request with a spreadsheet. The visible tabs had been deleted, but one stayed out of view, marked only by three dots on the tab bar, and the quality check looked at the response tab alone. The hidden tab listed &lt;a href=&quot;https://www.psni.police.uk/sites/default/files/2023-12/Protecting%20from%20Within%20a%20review%20of%20the%20PSNI%20data%20breach%208th%20August%202023.pdf&quot;&gt;all 9,483 officers and staff&lt;/a&gt;, with surname, rank and location, and it stayed on a public website for almost two and a half hours. The service later &lt;a href=&quot;https://www.bbc.co.uk/news/articles/cg4q26vnrklo&quot;&gt;failed to have its £750,000 fine reduced&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;In February 2014 Australia’s immigration department published a statistics report as a Word file with &lt;a href=&quot;https://www.oaic.gov.au/privacy/privacy-assessments-and-decisions/privacy-decisions/Investigation-inquiry-reports/department-of-immigration-and-border-protection-own-motion-investigation-report#:~:text=embedded%20the%20Microsoft%20Excel%20spreadsheet&quot;&gt;the source Excel spreadsheet embedded in it&lt;/a&gt;: personal information on about 9,250 asylum seekers, on the department’s website for about eight and a half days. Department policy called for seven reviewers to clear such a report, in hard copy. The digital file was not checked for embedded data.&lt;/p&gt;
&lt;p&gt;We built BADR for the minute before such a file leaves. It is an MVP demo in its presentation stage, not a product yet. It reads a Word, Excel, PowerPoint or PDF file the way a recipient’s software could, lists what the file gives away with the real values found, and returns a clean copy. At &lt;a href=&quot;https://albadr.ai&quot;&gt;albadr.ai&lt;/a&gt; the scan and the clean copy are free, one file at a time, up to 10 MB. Masking identifiers inside the document, the optional AI explanation and the API need a key. It reads text only: images and scanned pages are not read.&lt;/p&gt;
&lt;h2 id=&quot;office-software-hides-rather-than-deletes&quot;&gt;Office software hides rather than deletes&lt;/h2&gt;
&lt;p&gt;The people in these stories used their software as it was designed. Hiding a sheet, deleting the sheet a pivot table was built on, pasting a chart into a slide: none of these takes the content out of the file. Our own reports say it in the formats’ terms. On a workbook with a pivot table the report says: “A copy of the pivot table’s source data stays in the file, even if the source sheet was deleted.” A PDF that was saved several times gets this line: “Each incremental save appends to the file without erasing: earlier versions of text, fields or annotations remain readable.” And a slide deck with a chart in it: “The whole Excel workbook behind the chart travels with the deck, other sheets included.”&lt;/p&gt;
&lt;p&gt;None of this is news to the people who write guidance. The NSA’s guide &lt;a href=&quot;https://fas.org/publication/nsa_redacting_with_confidence/&quot;&gt;Redacting with Confidence&lt;/a&gt;, first issued in December 2005, says “&lt;a href=&quot;https://ftp.mirrorservice.org/sites/ftp.wiretapped.net/pub/security/info/papers/operating-sys/nsa-guides/other/how-to-publish-sanitized-reports-converted-from-msword-to-pdf.pdf&quot;&gt;The most common mistake is covering text with black&lt;/a&gt;” and, of anything that must not be read, “The item must be deleted.” A fact sheet from Canada’s federal privacy commissioner, dated 2006 and now archived, lists &lt;a href=&quot;https://www.priv.gc.ca/en/privacy-topics/technology/02_05_d_30/#:~:text=Track%20Changes&quot;&gt;tracked changes, comments, the names of current and previous authors, revisions, versions and hidden text&lt;/a&gt; among what travels inside a shared document. Microsoft’s own support page says its Document Inspector &lt;a href=&quot;https://support.microsoft.com/en-us/office/remove-hidden-data-and-personal-information-by-inspecting-documents-presentations-or-workbooks-356b7b5d-77af-44fe-a07f-9aa4d085966f&quot;&gt;cannot detect text hidden as white on a white background&lt;/a&gt;, and that in Excel it reports but cannot remove embedded files, macros and cached PivotTable data. The page describes a manual check of the one file that is open.&lt;/p&gt;
&lt;p&gt;Knowing has not been enough. Researchers at Inria examined &lt;a href=&quot;https://arxiv.org/abs/2103.02707&quot;&gt;39,664 PDF files published by 75 security agencies in 47 countries&lt;/a&gt;, and 33% of them revealed the identity of the person who created the file. Of all those files, 24% had been cleaned before publication, and for 65% of the cleaned ones the cleaning was weak: hidden information could still be recovered. Of the 82 PDFs the study took from the NSA’s own website, 45 still carried their full metadata.&lt;/p&gt;
&lt;p&gt;Quebec’s Commission d’accès à l’information received &lt;a href=&quot;https://www.cai.gouv.qc.ca/uploads/pdfs/CAI_RAAG_2023-2024.pdf&quot;&gt;444 incident notices in 2023-2024&lt;/a&gt;, and accidental communication was among the causes in 58 of them. Under the province’s private-sector privacy act, as amended by Law 25, a court can &lt;a href=&quot;https://www.cai.gouv.qc.ca/protection-renseignements-personnels/information-entreprises-privees/sanctions-entreprises-poursuites#:~:text=Montant%20des%20amendes&quot;&gt;fine an enterprise up to $25 million or 4% of its worldwide turnover&lt;/a&gt;, whichever is greater.&lt;/p&gt;
&lt;p&gt;Files now go to AI assistants as well. In one security vendor’s monitoring data, drawn from its own customers and so not a random sample, &lt;a href=&quot;https://www.harmonic.security/resources/genai-in-the-enterprise-its-getting-personal#:~:text=26.38%25&quot;&gt;26.38% of the files uploaded to generative-AI tools&lt;/a&gt; between July and September 2025 contained sensitive information, up from 22% the quarter before. Some cases are on the public record. In New South Wales a former temporary staff member &lt;a href=&quot;https://www.nsw.gov.au/departments-and-agencies/nsw-reconstruction-authority/northern-region/northern-rivers/resilient-homes-program-data-breach#:~:text=10%20columns%20and%20more%20than%2012%2C000%20rows&quot;&gt;uploaded an Excel spreadsheet of 10 columns and more than 12,000 rows to ChatGPT&lt;/a&gt; in March 2025, exposing personal information of 2,031 people in a flood-recovery housing program.&lt;/p&gt;
&lt;h2 id=&quot;what-badr-does&quot;&gt;What BADR does&lt;/h2&gt;
&lt;p&gt;You drop one file on the page. A report comes back: one verdict line with a risk level, then what the file gives away, grouped by kind and sorted by risk. Each clean-up option is a switch, and one click returns the clean copy. On the site’s own sample, a fictional offer in Word, the verdict begins “This file gives away 27 items, 10 of them high risk”.&lt;/p&gt;
&lt;p&gt;It looks in places nobody opens by hand: the saved name of a network printer, a header that prints the file path, a data connection that carries its server, its account and sometimes its password, the workbook behind a chart, the earlier versions an edited PDF still holds, an identifier that appears nowhere on the page because only a formula or a link’s address holds it. It recognizes 11 kinds of identifier, among them social insurance numbers, Quebec health insurance numbers, payment cards and IBANs.&lt;/p&gt;
&lt;h2 id=&quot;how-it-is-put-together&quot;&gt;How it is put together&lt;/h2&gt;
&lt;p&gt;These are its properties, each with its cost. The page is interface only. The engine runs on our server and never reaches a browser, which receives about 100 KB of script (36 KB compressed), and that script calls its own origin and nothing else: no tracking, no advertising, no analytics. The cost is that a document has to travel to us, over an encrypted connection.&lt;/p&gt;
&lt;p&gt;So a file is read in memory and never kept. The server keeps nothing of a document between two requests, and when you ask for the clean copy the page sends the document again. The logs hold counts and sizes (the kind of file, its size, how many findings, how long it took) and no file name, content or address. The file is handled in the data centre that receives it, generally the nearest, which may be outside Quebec or Canada. The cost is every feature that would need a memory of your documents. What it buys is that there is no store of documents to protect or to hand over.&lt;/p&gt;
&lt;p&gt;The engine is self-contained. While it reads or cleans it calls nothing outside itself, no AI service included, and it never executes anything a file contains, whether macro, script, field or formula. The same file always gives the same report, apart from the line that says how long the scan took. The cost is that detection works by patterns, which can miss something or flag too much, so every finding is shown with its location for a person to judge. When a part of an Office file cannot be read, the report says so, the verdict is never “all clear”, and no clean copy is made.&lt;/p&gt;
&lt;p&gt;All of it runs under a 128 MB memory cap, the allowance of a small sandbox on an edge platform. The hosted demo’s limits are sized for that sandbox: 10 MB per document, 20 documents and 25 MB per request. Each request also has a ceiling on the work it takes on. A document that does not fit is listed in the answer with the reason, and the client sends it again in a smaller batch.&lt;/p&gt;
&lt;p&gt;The clean copy holds what you chose to keep. Deleted text, earlier versions and content left behind by past edits are not in it. Masking removes the characters, where a black box would only sit on top of them, and a masked copy is read once more so that the report can count anything that could not be masked. Hidden sheets, rows and columns are made visible rather than deleted, so that you see what the recipient could have found.&lt;/p&gt;
&lt;h2 id=&quot;what-we-measured&quot;&gt;What we measured&lt;/h2&gt;
&lt;p&gt;The first table times the scan alone, the engine reading a file and writing its report, on one desktop processor from 2012 (an Intel Core i7-3770). Each value is the median of 3 runs after a warm-up, except that a file whose scan and clean copy together take more than 20 s is timed once. The files are synthetic, built at real sizes from text that varies like prose. “With photos” means 10% text, and random bytes standing in for the pictures.&lt;/p&gt;
&lt;div class=&quot;table&quot;&gt;&lt;table&gt;
&lt;thead&gt;&lt;tr&gt;&lt;th&gt;Scan, engine only&lt;/th&gt;&lt;th class=&quot;to-end&quot;&gt;0.25 MB&lt;/th&gt;&lt;th class=&quot;to-end&quot;&gt;1 MB&lt;/th&gt;&lt;th class=&quot;to-end&quot;&gt;5 MB&lt;/th&gt;&lt;th class=&quot;to-end&quot;&gt;10 MB&lt;/th&gt;&lt;/tr&gt;&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;&lt;td&gt;Word, text only&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;107 ms&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;573 ms&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;3.5 s&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;not measured&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;Excel&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;109 ms&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;610 ms&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;4.5 s&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;not measured&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;PDF, text only&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;106 ms&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;318 ms&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;405 ms&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;540 ms&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;Word with photos&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;12 ms&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;44 ms&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;214 ms&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;435 ms&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;PowerPoint with photos&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;12 ms&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;36 ms&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;187 ms&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;410 ms&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;PDF with photos&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;3 ms&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;7 ms&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;31 ms&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;60 ms&lt;/td&gt;&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;
&lt;p&gt;Up to 1 MB, every kind of file is read in at most 0.61 s. At 25 MB, a Word or PowerPoint file that is mostly photos takes 1.3 s, and a PDF 148 ms. Three cautions go with the table. Text compresses well, so the text-only Office files are denser than their size suggests: unpacked, the 1 MB Word file holds about 7.8 MB of text and markup and the 1 MB Excel file about 5.5 MB, and the hosted demo refuses an Office document with more than 5 MB of it. From 1 MB up, those two rows are the engine’s speed and not something the demo will do for you. Nor are the 25 MB figures, since the demo stops at 10 MB. And the text-only PDF row flattens after 1 MB because in a long PDF the scan reads the text of at most the first 400 pages, and says so in its report.&lt;/p&gt;
&lt;p&gt;The clean copy costs more than the scan. With every option on and identifiers blacked out, a text-heavy file took 2.4 to 4.1 s at 1 MB, and a 5 MB spreadsheet took 81 s. These too are the engine’s figures, for files denser than the demo accepts. A 1 MB file that is mostly photos took 0.03 to 0.21 s.&lt;/p&gt;
&lt;p&gt;The second table is the worst case through the API: every option, identifier masking, a ZIP coming back. Each scenario ran once with the size limits lifted, then again under a 128 MB memory cap set on a server process to stand in for the edge sandbox.&lt;/p&gt;
&lt;div class=&quot;table&quot;&gt;&lt;table&gt;
&lt;thead&gt;&lt;tr&gt;&lt;th&gt;Worst case, one run each&lt;/th&gt;&lt;th class=&quot;to-end&quot;&gt;Time&lt;/th&gt;&lt;th class=&quot;to-end&quot;&gt;Under the 128 MB cap&lt;/th&gt;&lt;/tr&gt;&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;&lt;td&gt;One Word, PowerPoint or PDF document, 1 MB&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;0.05 to 0.14 s&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;fits&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;One PDF, 15 MB&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;0.8 s&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;fits, close to the cap&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;20 mixed documents of 1 MB&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;5.2 s&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;6.0 s&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;40 mixed documents of 1 MB&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;10.3 s&lt;/td&gt;&lt;td class=&quot;to-end&quot;&gt;11.8 s&lt;/td&gt;&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;
&lt;p&gt;These figures leave things out. They time the engine, and what you wait for adds the upload and the answer’s way back. Nothing was measured on the edge platform’s own processors, and every file was synthetic. And nothing here measures accuracy: we have no precision or recall figure to publish yet. The web edition carries 189 automated tests, which show that the engine does what we specified and cannot show that what we specified catches everything.&lt;/p&gt;
&lt;h2 id=&quot;why-we-call-it-an-ai-solution&quot;&gt;Why we call it an AI solution&lt;/h2&gt;
&lt;p&gt;Two kinds of AI are in BADR, and we keep them apart.&lt;/p&gt;
&lt;p&gt;The engine is the older kind, knowledge written down as rules, and it is built with the methods of data science. Each of its steps has an ordinary name: entity recognition for finding a social insurance number in running text, validation for confirming it by its check digit, structural analysis for reading a file’s structure and not only its text, rule-based classification for ranking what was found. None of this is our invention. NIST’s report on de-identification, surveying work on medical records, names &lt;a href=&quot;https://nvlpubs.nist.gov/nistpubs/ir/2015/NIST.IR.8053.pdf&quot;&gt;rule-based systems as one of the two primary techniques explored&lt;/a&gt; for finding identifiers in free text. Nothing in the engine is machine-learned or statistical, and that is deliberate: every finding traces back to a rule a person wrote, and nothing is learned from your file. It is why the FAQ on our site can say that no AI reads your documents. In the sense most people mean in 2026, a language model reading the file, none does.&lt;/p&gt;
&lt;p&gt;Whether the older kind still counts as AI depends on who is asked. Stanford’s One Hundred Year Study on Artificial Intelligence calls it &lt;a href=&quot;https://ai100.stanford.edu/2016-report/section-i-what-artificial-intelligence/defining-ai#:~:text=odd%20paradox&quot;&gt;the “AI effect”&lt;/a&gt;: AI brings a technology into common use, people get used to it, and it stops being considered AI. Official texts draw the line in different places. Recital 12 of the &lt;a href=&quot;https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=OJ:L_202401689&quot;&gt;EU AI Act&lt;/a&gt; counts logic- and knowledge-based approaches among AI techniques and, in the same recital, says the definition “should not cover systems that are based on the rules defined solely by natural persons to automatically execute operations”. Canada’s federal &lt;a href=&quot;https://www.tbs-sct.canada.ca/pol/doc-eng.aspx?id=32592#:~:text=rules%2Dbased%20systems&quot;&gt;directive on automated decisions&lt;/a&gt;, which governs administrative decisions in the federal government, steps around the question: its definition of an automated decision system covers any technology that assists or replaces the judgment of a human decision maker, and names rules-based systems alongside machine learning and neural networks. We make no claim about how a regulator would class our engine.&lt;/p&gt;
&lt;p&gt;The second kind is the one most people now mean by the word: a language model. It is optional. For key holders, one click makes one call that explains the findings in a few plain sentences. It writes words only, and the risk level stays the engine’s. It receives the findings and never the file, after they have been de-identified by ordinary data-science means: data minimization, pseudonymization (stand-ins for detected identifiers and for the names of the file’s authors and editors, with the real values put back only in your browser), generalization and truncation. We do not call that anonymization, because an excerpt “can still hold another person’s name, an amount or a sentence from the document”. On our hosted demo that call goes to an outside AI service provider. In a deployment of the customer’s own, the model can be one of the customer’s choosing, including one that runs on its own machines. Without it the standard report is complete.&lt;/p&gt;
&lt;h2 id=&quot;where-it-fits-and-how-a-client-connects-it&quot;&gt;Where it fits and how a client connects it&lt;/h2&gt;
&lt;p&gt;The step is the same wherever a document crosses a boundary: a law firm sending a contract whose tracked changes still hold the text it deleted, a public body answering an access-to-information request with a spreadsheet, someone about to upload a report to an AI assistant, a pipeline that publishes or archives documents that nobody reads one by one.&lt;/p&gt;
&lt;p&gt;Integration is one HTTPS call with a multipart upload, from any language or from a workflow tool’s HTTP step. A call carries up to 20 documents. One document comes back as the cleaned file. Several come back as a ZIP with a manifest that says, for each, which options were applied and what was removed, for the client’s own logs. A document that fails is listed in the manifest while the others still come back, and a refusal over a limit states what was measured, the limit and why the limit exists.&lt;/p&gt;
&lt;p&gt;We ship no connectors. The call could sit where your documents already pass: the outbound step of a mail gateway, the publish step of a document management system, a script that watches a shared folder, the step before an upload to an AI assistant. We host the API today. Running it in a client’s own cloud account or on its own servers is available on request, and is not packaged yet.&lt;/p&gt;
&lt;h2 id=&quot;moving-it-to-another-host&quot;&gt;Moving it to another host&lt;/h2&gt;
&lt;p&gt;The limits are settings, not code: on a larger host the number of documents, their size and the ceiling on work are raised by configuration. The engine already runs on two kinds of host, the edge sandbox in production and an ordinary server process in development and in every test above.&lt;/p&gt;
&lt;p&gt;The whole engine also exists a second time, written for a second runtime as a native Windows application that works with no server at all: no file leaves the computer. It is not released. With the first engine there to check against, the second took a single day, 7 November 2026. A corpus of 48 documents went through both in five modes: every scan identical, every clean copy equivalent. The one difference found was a few bytes in a finding that prints the size of an embedded archive. The application’s installer is 2.1 MB.&lt;/p&gt;
&lt;h2 id=&quot;what-the-demo-does-not-do-yet&quot;&gt;What the demo does not do yet&lt;/h2&gt;
&lt;p&gt;BADR is a demo, and these are its limits today. It reads text only: text inside a picture or a scanned page is neither checked nor masked. In a long PDF the text of at most the first 400 pages is read, and the report says so. Word, Excel and PowerPoint 97-2003 files are not supported. The identifiers are tuned for Canada and Quebec. Detection works by patterns, so a valid-looking number that is not an identifier can be flagged and a malformed one can be missed. A file can be 10 MB at most, and an Office document can hold at most 5 MB of text and markup once unpacked. An encrypted PDF is analyzed for its structure only and gets no clean copy. It has had no outside security audit, and it carries no service-level commitment. Check every clean copy before you send it; in an unattended step, that check is a person or a rule reading the manifest.&lt;/p&gt;
&lt;p&gt;The scan is free at &lt;a href=&quot;https://albadr.ai&quot;&gt;albadr.ai&lt;/a&gt;. No price is posted. For a key, or to talk about running BADR inside your own systems or acquiring it as it stands, write to contact@albadr.ai.&lt;/p&gt;</content>
</entry>
</feed>
