Ontelic builds Vellit — an AI that helps people know their rights, and helps lawyers and advocates work faster. Tuned hard against hallucinations, starting in Uzbekistan and expanding across the CIS toward the EU.
Ontelic isn't just another startup wrapping someone else's API. We build Vellit to help people know their rights — and to help lawyers and advocates work faster, grounded in real law, not guesses.
Ontelic was founded with a single goal: to make the law legible. Legal knowledge has always belonged to those who could afford a lawyer. Vellit exists to close that gap — a system that reads legislation the way it actually is, cites where an answer comes from, and says "I don't know" when it doesn't.
Vellit is fine-tuned and constrained specifically to minimize hallucinations in a domain where a wrong answer has real consequences. Every response is traceable to a source — either the live text of the law or a dated, versioned snapshot of it. We never let the model blend the two silently.
Our philosophy is simple: the model must be predictable, safe, and useful. Legal answers go through source-grounding and explicit uncertainty handling before they reach a user. We don't chase hype — we chase accuracy.
We don't run our own legal database from scratch. Vellit is built on top of official public sources — live government portals and open legislative datasets — transformed into a system that actually understands the difference between what the law says today and what it said in an archived snapshot. This is real engineering, aimed at one very concrete problem: people not knowing what they're entitled to.
Updates on the company and on Vellit as it moves through closed testing.
Vellit 1.0 is fast and cheap for everyday questions. Vellit 1.2 handles hard cases and case dossiers.
Law search, memory and dossiers stay off until a question needs an exact citation. Everyday chat stays fast and costs less.
We will show Vellit live and meet the local tech and legal community.
A limited group of testers in Uzbekistan is using Vellit. No public release date yet.
Ontelic is discontinuing the Orynt model line and the Leili image generator to focus entirely on Vellit.
Heading to Google Play first, with an App Store release to follow.
One assistant that separates live legislation from archived snapshots, asks follow-up questions on complex cases, and never guesses when it isn't sure.
Most questions don't need a law search. Vellit starts light and fast, and turns the heavier tools on for the exact moment a citation matters.
Only two tools are visible to the model at the start: a chat title and a switch. No search, no memory. Cheap and fast for everyday questions.
When a question needs an exact citation, the model asks to switch it on. You see it happen, with the reason.
Law search, articles, memory and case dossiers unlock. It reads the source and tells you whether it is the live text or a dated snapshot.
It answers with the source named. When the question is closed, the protocol switches off and the chat goes light again.
Built for everyday and mid-complexity questions. Cheap and quick, so it can think at length and still stay affordable.
Built for hard cases. Runs case dossiers, follows a long investigation and catches inconsistencies. It costs more, which is why it is a separate choice.
Ask about your situation in plain language. Every answer names its source: the live text of the law, or a dated snapshot. If the two disagree, Vellit says so. It does not average them into a guess.
When it isn't confident, it says so and points you to what to check next, instead of inventing a plausible-sounding article number.
A long-running investigation into your specific case. Vellit can pause to ask a clarifying question, remembers everything gathered so far, and picks up where it left off. Nothing is lost, nothing is rushed.
When the investigation concludes, you get a clear report to download as a PDF or share by private link, useful when you need to hand your situation to a lawyer, an employer or a government office.
Two sources, never silently blended: the live government legal portal, and a versioned, dated open dataset of Uzbek legislation. Vellit always tells you which one an answer came from.
If a lookup fails, Vellit says so. A failed request is never returned as a quiet "nothing found". You always know whether it found nothing, or simply couldn't check.
Not every case starts from a blank page. A short list of common situations — a rental agreement to check, a dismissal to contest, a traffic fine to appeal, a consumer complaint — starts a dossier already pointed at the right law, with your own details added on top.
Any article can be explained in a few sentences of ordinary language instead of legal phrasing: what it regulates, what it means for you in practice. Explained once per article and language, then served instantly after that.
For a dossier, Vellit can lay out the case as a graph: facts, evidence and claims linked to the articles that apply to them, each connection marked as supporting or contradicting. It's a map of why the conclusion holds, not just the conclusion itself.
Vellit doesn't have one big prompt that tries to do everything. It reaches for a specific capability only when a question actually needs it — nothing runs by default.
Vellit can search the live government legal portal directly, or fall back to a dated open snapshot of the same legislation when the portal is unreachable. It reads either a specific article or a whole act, and it always prefers the live source for a final answer.
Mid-case, Vellit can stop to ask exactly one clarifying question, and save what it's found so far before doing it. Come back later and nothing gathered along the way is lost.
Durable facts about your situation — your line of work, a case you're already in the middle of — can be kept across conversations, and removed just as easily if you'd rather it forgot.
When a question genuinely depends on where you are — local jurisdiction, a regional office — Vellit asks for your location through the app's own permission prompt. Never guessed, never inferred from chat text alone.
Opening a full case dossier or laying out a connection graph is always offered first, as something you accept or decline — not something that starts the moment Vellit thinks it might help.
Legal systems don't generalize easily. Each jurisdiction has its own codes, language and institutions, so we are not expanding by translation. We rebuild the grounding for each country before we call it live.
Ontelic's first attempt: a general-purpose model line and an image generator. Retired in August 2026 because of small demand and a market full of stronger models from far larger labs. That experience pointed us to a narrower, underserved problem.
Running with a limited group of testers. No fixed public date. Vellit launches fully self-sufficient from version one: legal chat, multi-step case investigations and shareable reports, all from day one.
Ontelic will be at ICTWEEK, showing Vellit live and meeting the local tech and legal community.
Investment talks with Imkon Ventures, growing out of the same ICTWEEK week in Tashkent. If it closes, it funds a faster path to public launch — the roadmap below doesn't change, it just moves sooner.
Vellit's first launch. Full coverage of lex.uz as the live source, plus an open, dated dataset of legislation from 1991 to present as an archival fallback. This is where we prove the models, the sourcing discipline and the dossier workflow.
Country by country, starting with jurisdictions close to Uzbek civil and administrative law. Each gets its own live-source integration and its own dated dataset, never a shared, blurred model of "post-Soviet law in general".
EU law spans supranational regulation and national implementation at once. This stage starts only once the sourcing and hallucination-minimization discipline from stages 1 and 2 has been proven at scale.
What doesn't change at any stop: live law and archived snapshots are never mixed silently, every answer is sourced, and the model says "I don't know" rather than invent an article number. That discipline is the product. The country list is just where we apply it.
Ontelic is small on purpose. Two people, no committee, no layer of process between an idea and its deployment.
Nat Drexov — founder and sole developer of Vellit. Years of self-study in programming and systems design led him to a single, specific problem worth solving: people don't know their own rights, and the tools that promise to explain the law can't be trusted to get it right.
Without formal education in computer science, Nat learned to program through books and forums, and gradually came to an understanding: the most useful place for AI isn't answering trivia — it's standing next to a person at the exact moment they need to know what they're legally entitled to, and telling them the truth, sourced and dated.
That conviction shaped Vellit's core design rule: never blend live law with an archived snapshot without saying so. Nat believes a legal assistant's first duty isn't to sound confident — it's to be honest about what it actually knows and where that knowledge came from.
Today Nat leads all directions of the company — from the sourcing architecture to the product itself. His approach to development is characterized by an unwavering commitment to minimalism: remove everything superfluous, leave only what is essential. This philosophy shows up in Vellit's product design as much as in its technical implementation.
Every decision passes through that lens. There is no room for a shortcut that trades accuracy for speed. If a feature doesn't bring direct benefit to someone trying to understand their rights, it's removed. That takes more time to build — but the result is a system that's predictable, safe, and honest about its own limits.
A legal AI that sounds certain is more dangerous than one that admits doubt. We didn't build Vellit to have an answer for everything. We built it to know the difference between what it knows and what it doesn't — and to say so, every time.Nat Drexov CEO Ontelic
The pullquote, printed. Not a mockup — this is on an actual shirt.
Ramazan — DevOps engineer keeping Vellit's backend alive on a shared FreeBSD box that was never designed for this. He was a friend of Nat's for years before he was a colleague, which is exactly why he's here.
Ramazan's job is unglamorous by design: deploys that don't break the running service, webhooks that stay registered, a JWT layer that doesn't quietly rot. When something in the pipeline silently pushes a stale build, he's the one who notices before it becomes a user-facing bug. Reliability isn't a feature he ships — it's the absence of things going wrong that nobody sees.
He and Nat go back to before Ontelic existed. Long before there was a company, a product, or a plan — there was just two people who trusted each other, in a city where trust is worth more than a resume.
That trust showed up in the most literal way possible early on: at a point when Nat didn't have a laptop to build on and Vellit was still nothing but an idea in his head, Ramazan bought him one and paid for it in crypto — no paperwork, no interest, no expectation of anything back. It's the kind of thing that doesn't show up on an org chart, but it's part of why this company exists at all.
Today that same instinct — show up, fix the thing, don't make a production out of it — is what keeps Vellit's infrastructure quiet and boring in exactly the way infrastructure should be.
"Before he ever touched a config file for us, Ramazan sent me a laptop I couldn't afford — paid in crypto, no strings. Every deploy he ships since is just the same thing, at a different scale." — Nat Drexov, on Ramazan
Vellit is more than a single model. The discipline around sourcing, cost and error handling is what keeps it honest.
Every response is tagged with where it came from: the live government legal portal, or a dated, versioned open dataset. The two are never blended without saying so.
A failed lookup is always returned as an explicit error, never as a quiet "no results". You always know whether Vellit found nothing, or simply couldn't check.
Search, memory and dossiers are off until a question needs them. You pay for heavy work only when it actually happens, and you can see when it starts.
Live lookups against the government legal portal honor its rate limits by design. Vellit was built to be a good citizen of the sources it depends on.
Personal data is encrypted at the application level, not just on disk. The encryption key lives outside the source code, in the environment or an external key store, never checked into the codebase.
Uzbek-language legal text is translated to Russian and English by a model that runs on Vellit's own server, not by sending your questions to a third-party translation service.
If the translator, Telegram or the PDF engine has a bad day, that one feature degrades. It doesn't take the rest of the API down with it.
© Ontelic. Vellit is fine-tuned and constrained for legal accuracy: long chat sessions, multi-step case investigations that can pause on a clarifying question, and a strict separation between live and archived law. In a legal context, a confident wrong answer is worse than no answer at all.
Important: Vellit is a tool for understanding your rights, not a substitute for a licensed lawyer in matters that require one. We are upfront about that distinction, every time it matters.No subscription tier to pick blind. Everyday chat is billed by actual tokens, a case dossier is a flat price per session, and a small free daily budget covers casual use without a card on file.
Top-ups are bought as Telegram Stars — no local card processor, no separate merchant registration, and no crypto wallet to manage. A star purchase converts to your balance once, and every later charge is metered in that same balance, never re-converted or re-priced after the fact.
Running low never cuts you off mid-answer: spending only ever happens on an action you took, and you're told before you're blocked, not after.The API doesn't trust a request until it proves who it is. A handful of paths are public on purpose — everything else needs a valid session, and every session only ever sees its own data.
Access is refused unless a path is explicitly listed as public. A new endpoint that forgets to declare who can see it is closed by default, not open to everyone by accident.
Every request against a specific case is checked against who created it. A mismatch returns "not found", not "forbidden" — so guessing session IDs can't even confirm one exists.
Signed-in users are limited per account, not per network address — because many real people can share one address on a mobile carrier. Anonymous requests are limited by address instead, and OTP requests get their own, tighter limit, since every one of those costs real money to send.
Names, phone numbers and case details are encrypted before they ever touch disk, with the key held outside the source code — in the environment, or an external key store, never checked into the codebase.
A crash returns a generic message and a request id, not a stack trace — the detail goes to the server's own log, where it's actually useful, instead of into a response anyone could read.
A legal answer is often only useful once it leaves the app — handed to a lawyer, an employer, or an office window. Vellit is built with that handoff in mind.
No password to invent or forget. A code arrives by SMS, by voice call, or as a contact confirmation inside Vellit's own Telegram bot — whichever actually reaches you.
When a dossier concludes, it isn't just left sitting in the chat. It becomes a structured report you can read as Markdown, download as a PDF, or send as a private link — built for the moment you need to hand your situation to someone else.
Pending dossier notifications live on the server against your account, not the device — reinstall the app, switch phones, come back a week later: nothing that finished while you were away goes missing.