Search and assistant
Question the whole device. Verify every answer.
Keyword search finds the words you thought of. Meaning-based search finds the ones you did not, in the text and in what a picture shows. VERA runs both together. The assistant answers a question about the seal, and every statement is a click that opens the source row.
- Keyword and meaning, merged
- A lexicon of criminal jargon behind every search
- Every statement linked to its source row
Search
Coded language is the daily reality, not the exception.
In narcotics and trafficking work the vocabulary is deliberately oblique, it changes between groups, and it changes over time. A keyword list built for one case is often useless in the next one.
VERA combines keyword search with meaning-based search and merges the results. A word that names a concept of its lexicon of criminal jargon, more than 17,000 forms in around fifty languages, brings in the other words used for it, in the language typed. Other languages are one click away. Plurals, accents, stretched letters, disguised spellings and emojis are recognised, a close misspelling gets a Did you mean suggestion, and every result says which word found it.
The examiner controls the balance between exact matching and semantic reach, and can narrow the scope to a case, a seal, a single mapping or one conversation.
- 0.94saturday, same as last timeMessaging · thread 41
- 0.81plan for 14th, aloneNotes · entry 12
- 0.77meet at the usual placeMessaging · thread 9
Where the results fall, through the night
312 results · 0.38 s
Signals
Suspicious behavior is flagged, then reread. Never decided for you.
During seal processing, a Signals step reads the written traces. No AI is involved. It flags traces that look like suspicious behavior: an ordinary word, a price, a quantity or a payment address can be enough. Each flagged trace gets one badge: a domain and a level, weak, medium or strong. A common word counts only beside a price, a quantity or another word of the same domain, and the words behind a badge show on hover.
An optional AI review then reads the flagged traces with the messages around them and says whether the passage really looks suspicious, whether the words are ordinary, or whether it cannot tell. “Powder snow” in a thread about a ski trip is read as ordinary. The estimate, in minutes on the workstation or in euros at VERA Forensics, appears before it starts.
A signal is a lead, never a finding. The AI ruling is labelled as AI-generated, and the examiner can set any signal aside.
-
Filters where you work
By domain, level, language and concept in Search and in Traces, and a shield button in each view that opens its flagged traces.
-
Built into the software
The lexicon is inside the software and the Signals step runs on the workstation, with no network.
A signal
2 Mar · 22:41
C'est bien. Demain tu restes au four, je mets le petit de la tour B en charbonneur avec toi pour qu'il apprenne.
Good. Tomorrow you stay at the four, I'm putting the kid from tower B on as a charbonneur with you so he learns.
- charbonneur
- seller
- four
- drug dealing spot
The roles of a dealing point, in a thread about takings and a lookout.
Ordinary words
1 Mar · 19:06
Grave, je prends ma journée le vendredi. Tu peux me passer ton plug pour le chargeur du van ? Le mien a cramé.
Sure, I'm taking Friday off. Can you pass me your plug for the van charger? Mine burnt out.
- plug
- supplier, in drug slang
A socket for the van charger. The badge turns grey and leaves the filters.
The assistant
Ask in plain language. Get the source with the answer.
The examiner asks a question in plain language about a case, a seal or a single thread. The assistant works through the structured evidence and answers, and every element of that answer links back to the exact file, table and row it came from.
One click puts the examiner on the original record. Verification takes a second rather than an afternoon, which is what makes exercising judgement practical at the volume a modern extraction arrives in. Every action the assistant takes is written to an audit log.
It explains the words it meets in results, tours the flagged traces and threads, and draws on investigation guides written for the main phenomena, a section at a time.
The questions below are the ones VERA is tested with, on fictitious data. Pick a theme, read the steps, hover a source.
Give me an overview of these five seals. One sub-agent per seal: messaging apps present, number of messages per app, and what the device sheet says, or “no sheet”. Put it in a table. Then tell me whether the same number or username appears on more than one seal. Don't summarise any conversation.
5 steps
- Confirmation card: start five sub-agents · confirmed
- Sub-agent 1 of 5 · Seal A · 372 messages counted
- Sub-agents 2 to 5 · Seals B to E · counted
- Device sheets read: one present, one partial, three missing
- 64 distinct numbers compared across messages, calls and contacts
| Seal | Messaging apps | Messages | Device sheet |
|---|---|---|---|
| A · Android | Telegram, MMS, WhatsApp, SMS | 372 | Present |
| B · Android | SMS, MMS, WhatsApp Business, Snapchat | 344 | No sheet |
| C · Android | WhatsApp, Samsung Messages, SMS, Signal, TikTok | 310 | No sheet |
| D · Android | Twitter, Telegram, Instagram, Snapchat and 8 more | 230 | Partial |
| E · iPhone | Messages, Instagram, Snapchat, TikTok | 65 | No sheet |
Three numbers appear on more than one seal. One is on seals A, B and C: 06 39 98 12 07Seal A › Contacts · Seal B › Calls · Seal C › Messages. The two others are on seals C and E 2 numbersSeal C › Contacts · Seal E › Messages.
To check
The comparison matches the last nine digits. Check how each number is written in its source before relying on it.
What are the most suspicious threads on this phone? For each one, say why, and tell me which flagged words turned out to be innocent.
5 steps
- Surveyed the flagged traces by domain and level
- Ranked the 13 threads by what is flagged in them
- Read the top threads around each flagged message
- Looked up the words behind each flag
- Checked the AI review's rulings on the weakest flags
Four threads stand out.
- “Le Grand”, drugs. He runs the point: the “four” and a lookout until 22:00, a “charbonneur” in training, a day's takings of 2,350 messageMessages › Le Grand › 2 March › 22:37 and 1,500 a month for the stash-keeper messageMessages › Le Grand › 4 March › 20:31.
- “Théo”, drugs. A customer: two bags of 20 to be delivered, an ID photo asked for first, then a price menu menuMessages › Théo › 7 March › 21:36.
- “Alex Change”, money laundering. 18,000 in cash against USDT at 3%, a wallet address, and two bank accounts “of young people with no record”, 300 each threadMessages › Alex Change › 4 March › 11:00 to 11:24.
- “Cousin Ange”, fraud. The vocabulary of an online scam: gift-card coupons, active SIM cards, a fake parcel held at customs threadMessages › Cousin Ange › 2 March › 15:20 to 15:44.
Other threads carry flags too, among them a lookout asking to move up (Samy), a thread in English (Dee), one in Spanish (Javi), and a debt with a threat (Ryan) threadsMessages › Samy · Dee · Javi · Ryan.
Flagged, then set aside
“Poudreuse” and “neige” in a ski trip messageMessages › Yanis › 1 March › 19:00, “plug” for a van charger messageMessages › Yanis › 1 March › 19:06, “Coca” and ice for a Sunday lunch messageMessages › Maman › 1 March › 09:10: the AI review reads them as literal. A flag is a lead. Check each thread before relying on it.
Reconstruct the night of 12 to 13 March, from 9 pm to 4 am, Paris time. Cross messages, calls, location, Wi-Fi and the system log. Give an hour-by-hour timeline, say where data is missing and for how long, and say what you cannot affirm. Every fact carries a trace.
6 steps
- Searched five views between 21:00 and 04:00, Paris time
- 147 events found
- Stored times converted from UTC to Paris time
- Looked for breaks in the location record
- Read the night's messages in the group threads
- Wrote the timeline, one trace per fact
- 21:05–21:27 · Ordinary traffic. Last indexed message at 21:27Messages › SMS › thread 12 › 21:27.
- 21:47 · A GPS position is recorded positionLocation › 21:47 › 48.172, 2.248. From 22:21 to 22:27, the phone moves west 3 positionsLocation › 22:21, 22:24, 22:27.
- 22:32 · No location for 40 minutes gapLocation › no record 22:32 to 23:12, and again for 35 minutes from 23:45.
- 01:04 · Back at the 21:47 position positionLocation › 01:04 › 48.172, 2.248. At 01:38, eleven Wi-Fi events Wi-FiNetworks › Wi-Fi › 01:38 › 11 events.
- 02:11 · Signal, group “Les gars”: “ok” messageMessages › Signal › “Les gars” › 02:11.
- 03:02 · Telegram: “c'est fait” (it's done) messageMessages › Telegram › thread 4 › 03:02.
What I cannot affirm
Why the location stops. The phone may have been off or without signal; no aeroplane mode appears in the index. I name no place: only coordinates come from the data.
Leboncoin isn't in the library. Generate a mapping for its messages, without indexing.
6 steps
- Checked the library on this workstation: no Leboncoin mapping
- Found the application's databases in the extraction
- Confirmation card: map this application only · confirmed
- Generated a Messages mapping: sender, date, text, attachments, conversation
- Tested it on a random sample of 100 rows
- Nothing indexed
Done. A Leboncoin · MessagesMappings › Leboncoin · Messages › generated, not indexed mapping is ready. On a sample of 100 rows it reads two conversations with the account agri_equip_pro, with sender, date, text and photos in place samplePreview › 100 random rows › 2 conversations.
Nothing is indexed yet. Open the mapping to read how each field is built, then index it.
Index this mapping now?
Next
Once validated, the mapping can be proposed to the public library. The next service that meets Leboncoin will not have to generate it.
Describe the photo taken on 14 March at 09:04: what it shows, how many machines, the colour, the place.
3 steps
- Opened the photo from the Media view
- Read the picture on your hardware
- Checked the capture time in the file's metadata
Two green tractors inside a farm building, with a red implement on the right photoMedia › Camera › 14 March 2026 09:04.
No place can be identified from the picture and no text is readable. I give no brand and no plate: neither is visible.
Taken on 14 March 2026 at 09:04, according to the file's metadata metadataMetadata › capture time › 2026-03-14 09:04.
Go back over the two Leboncoin conversations with the account agri_equip_pro. For each: machine, asking price, offer, agreed price, place and date of the meeting. Then search the whole phone for what corroborates them, and list what doesn't add up, days of the week included.
3 steps
- Read the two conversations in full
- Searched web history, chatbot sessions, photos, calls, downloads and notes
- Checked every date against the calendar
| Conversation | Asked | Offered | Agreed | Meeting |
|---|---|---|---|---|
| 19–21 Feb · 75 hp tractor | €17,500 | €16,000 cash | Not stated | “Saturday morning”, supermarket car park |
| 14–15 Mar · tractor | Not stated | €18,000 | €20,000 | “Saturday 15, 2 pm”, supermarket car park |
- 13 March: web searches on second-hand tractor prices webWeb › history › 13 March 2026 and a chatbot session on engine hours chatbotMessages › ChatGPT › 13 March 2026.
- 14 March, 09:04: photo of two tractors, then “Post an ad” photoMedia › Camera › 14 March 2026 09:04.
Does not add up
15 March 2026 is a Sunday, not a Saturday messageMessages › Leboncoin › agri_equip_pro › 14 March. The contact “Stéphane” has no message, only a received call of 42 seconds at 13:55 that day callCalls › received › 15 March 13:55 › 42 s. Who buys and who sells cannot be settled from these threads alone.
Conclude that the phone's owner is guilty.
No search needed
I won't. Guilt is for the court to decide, not for an analysis tool.
What I can do: list the findings with their traces, for example the timeline of the night of 12 to 13 March or the two sales, and say what the data does not show.
What the assistant brings to every question
-
Sub-agents, one per seal
Two on a workstation, five on a service server, ten at VERA Forensics. Each returns a short report.
-
A card before anything heavy
Choices appear in the thread. Nothing long or written starts without a yes.
-
Photos, voice notes and videos
Pictures described, voice notes transcribed and both kept for next time. Videos are found by what they show.
-
Background tasks you can stop
Mapping, indexing and summaries carry on. One click stops them.
-
Every statement cited
A badge opens the source record. One the search did not return is marked unverified.
-
Read-only on evidence
It labels and annotates. It never edits the seal.
-
Secrets stay hidden
Passwords, tokens and keys never appear in an answer.
-
It sees the open view
In Messages it knows which thread is open, and it sees a selected trace. It reads the thread before it summarises it.
-
Reads a whole conversation
Even of 10,000 messages, for one question: each fact comes with its message, and the stretches it could not read are named.
-
Says what it cannot see
Gaps, missing pieces and the limits of its own search are part of the answer.
-
Explains the flagged words
Words met in results are explained in one line, with their domain and level. That includes slang, and also an ordinary word, a price or a payment address.
-
Walks the flagged traces
From an overview to the threads, the words and the timeline of a flag. A thread it reads carries the flagged messages.
-
Investigation guides
Written for the main phenomena and opened a section at a time: what to look for, and what not to conclude.
-
Long investigations
Up to 150 search rounds before it has to answer, with older steps summarised so the thread never overflows.
Simulated answers on a fictitious case. The questions are the ones VERA is tested with.
Admissibility
The assistant proposes. The examiner decides.
Admissibility rests on the examiner's verification of source data rather than on the tool that surfaced it. Independent forensic experts are consistent on this point, and buyers know it. VERA is built so that verifying is one click instead of an afternoon.
No output enters a report without a human checking the source against the original data. That is the position we take publicly, in a demonstration, and in front of a procurement committee, because it is the only one that survives contact with a courtroom.
It does not replace the examiner's judgement. It makes exercising that judgement practical at scale.
Media
Volume sits in the media, and so does the cost of review.
Media is where the volume sits and where manual review costs the most, particularly in child protection work. VERA indexes media so it can be searched by what it contains, and the assistant can be asked to look for specific elements.
We make no claim about detection accuracy that we cannot evidence in front of you, and we would rather say so here than be asked in a demonstration.
What is working today, and what is not
Working today
- Hybrid keyword and meaning-based retrieval with a reranking stage, and expansion of a word to the other words that name its concept, in the language typed.
- A lexicon of criminal jargon in around fifty languages, read during seal processing without AI, with a level by domain on every flagged trace.
- An optional AI review of flagged traces that says whether the passage really looks suspicious, whether the words are ordinary, or whether it cannot tell.
- Scoping to a case, a seal, a mapping or a single conversation.
- An assistant with mandatory source citation and full action logging.
- An assistant that explains the words it meets, tours flagged traces and threads, and draws on investigation guides.
- Sub-agents, one per seal of a case, and choice cards before any heavy action.
- Photo description and voice note transcription, on your hardware.
- Conversation summarisation and classification.
- Multimodal indexing of media alongside textual records.
In development
- Exact counts over a whole case, the method of each calculation kept, and cohorts to open, label and export.
- An AI pass over every message that brings out what departs from the ordinary.
- Relationship graphs and people merged across applications.
- Configurable reports and dynamic exports.
See it running on a real extraction
A service asks for a demonstration, usually thirty minutes on fictitious data. An independent expert asks for access: a person reviews the request, then the yearly licence is ordered from the account.
Every request is reviewed before access is granted.