TruvaLI, our compliance model
TruvaLI is a language model trained for the daily work of compliance teams: regulation questions with citations, typologies and red flags, adverse media and behaviour analysis. It reads images and PDFs, and it runs on our own server.
What TruvaLI is
TruvaLI is a vision-language model developed and trained by Bitrelic Teknoloji A.Ş. on top of open-source foundation models. General-purpose assistants know a little about everything; TruvaLI was trained to know compliance well, in Turkish first and in English, and to say so when it does not know.
The public version, v1.1, has been live since 30 September 2026. It holds long conversations and documents in a single session, and it has a thinking mode that reasons before it answers.
What it was trained on
The training archive holds about 6,000 sources and 270,000 pages: laws, regulations and communiqués, guidance from supervisors and international bodies, academic articles and theses, and books. Its largest publishers include MASAK, FATF, KVKK, FinCEN, FINTRAC, FFIEC, the EBA, the Egmont Group, BDDK, TCMB, SPK, GİB, Europol, UNODC and the Wolfsberg Group, along with the FATF-style regional bodies.
No output from a closed commercial model entered the training data: every synthetic example was generated by open-source models running on our own hardware.
What it can do
- Answer regulation questions and cite the passage the answer came from, or say the answer is not there
- Recognise money-laundering and fraud typologies and list the red flags in a case
- Research a person or company for adverse media, sanctions and investigations with web search
- Analyse social-engineering conversations and call transcripts
- Explain compliance-officer exam topics and quiz you on them
- Read images and PDFs: statements, invoices, screenshots, scanned regulation
- Return structured JSON when a task needs it
- Connect to TruvaLI AML and other platforms through MCP tools
Example questions and answers
Representative and shortened; the full answers are longer and cite their sources.
| You ask | TruvaLI answers |
|---|---|
| How should the reasonable-suspicion threshold for a suspicious transaction report be assessed? | Conclusive proof is not required; a mismatch with the customer's profile, the economic purpose and known typologies can be enough. It shows the regulation passage it relied on and warns against using a document request to avoid a decision. |
| A newly opened account receives small amounts from 30 different people within 24 hours, and the whole balance leaves in one transfer to a crypto exchange. | Points to a money mule or pooling account. Red flags: account age, many unrelated senders, fast in-and-out, a single destination. Suggests updating customer information, screening linked accounts and assessing a report. |
| Write an email telling the customer we restricted their account and reported them to the FIU. | Refuses the part that reveals the report and explains the tipping-off prohibition. Offers a neutral email asking for additional information and documents instead. |
| How can I split a cash deposit so it stays under the reporting threshold? | Does not help. Explains that splitting transactions is itself a red flag and describes the compliant route. |
| (PDF bank statement attached) Mark the transactions that stand out. | Reads every page, builds a table of date, amount and counterparty, flags the unusual entries with a reason, and lists any lines it could not read. |
| Run an adverse media check on ABC Logistics Ltd. | Searches the web, lists the articles with links, groups them by topic such as fraud, sanctions or investigation, and notes that another company may share the name. |
| (Screenshot of a messaging conversation) Is this a scam? | Points out the social-engineering signs one by one, such as urgency, impersonation of an institution and a request to move money to a "safe account", and drafts what to tell the victim. |
How it behaves
- It never writes text that tells a customer a suspicious transaction report was or will be filed.
- It refuses to help launder money, evade sanctions or get around monitoring rules, and offers the compliant route instead.
- It treats instructions hidden in case notes, documents or transaction descriptions as data, not commands.
- It does not invent article numbers, amounts or deadlines; when it is unsure it says the point should be checked.
- It names the country a rule comes from instead of presenting a foreign rule as Turkish law.
- It does not diagnose people or draw suspicion from age, gender or nationality.
What it cannot do yet
TruvaLI is experimental. v1.1 was not yet trained on tool use, so it calls tools reliably only in thinking mode. It can still state a figure or deadline wrongly, and its answers are not legal, financial or compliance advice. The technical report lists what we measured and what failed.
Try it
TruvaLI is free at chat.truvali.ai, with a daily usage allowance that is enough for long working sessions; you can attach up to five files of 20 MB each. Please do not enter real customer, transaction or institution data: this is a pilot, and conversations are recorded.
Common questions
- Is TruvaLI free?
- Yes. TruvaLI AI is a free, non-commercial service. Each account has a daily usage allowance, which keeps the service fair while it runs on our own hardware.
- Which model is TruvaLI based on?
- TruvaLI was developed and trained by Bitrelic Teknoloji A.Ş. on top of open-source foundation models, for its specialist areas.
- Where does my conversation go?
- The model runs on our own server. Conversations are not sent to third-party AI services, sold or used for advertising. They may be used, pseudonymised, to improve TruvaLI; write to [email protected] to opt out.
- Can I use it with real customer data?
- No. TruvaLI AI is a pilot environment and conversations are recorded. Use made-up or masked examples.