Adylbek Z.

Teams bring me in when their data is scattered, unreliable, or locked in PDFs, APIs and Telegram channels that fight back, and they need it clean, trusted and usable by AI. I build pipelines that don't lose data quietly, and LLM layers whose output is checked, not assumed. Proven where mistakes are expensive: • Clinical AI live in 10+ clinics: 8 medical devices integrated, 95% diagnostic accuracy across 10,000 cases • Telegram data systems: channel discovery and monitoring over MTProto, 93 GB collected with full provenance • A 1M-record creator dataset built over a year, with an 85%-accurate prediction model on top Full case studies are in my portfolio below. WHAT I'M GOOD AT AI you can actually ship Anyone can call an API. The work is everything around the model: deterministic precomputation so it can't reason from numbers it misread, correction of invented output against a real registry, safety filters after the model, and an evaluation harness that proves accuracy instead of claiming it. Claude, Gemini, GPT, tool-calling, LLM-as-a-judge, guardrails, fine-tuning, scoring models. Systems where mistakes are expensive Medical records, money, personal data. Built so an error is loud instead of silent: audit trails, strict role-based permissions, encryption with key rotation, and adversarial review that hunts for what breaks before your users do. FHIR, ICD, LOINC, SNOMED CT, HL7, encryption, audit logging. Data nobody hands you Crawlers, scrapers, device protocols. Sources that fight back, rate limits, undocumented hardware. One system I built holds ~1M records across 1M+ scanned channels with a full year of history, plus a predictive model at 85% accuracy on top of it. That history can't be bought anywhere — it had to be collected before anyone needed it. Python, JavaScript, Airflow, ClickHouse, PostgreSQL, BLE, reverse engineering. HOW I WORK I take few projects at a time and I don't oversell. If a number of mine comes with a caveat, you'll hear it from me before you have to ask. Bring me the hard one — data you can't get at, a model whose output you don't trust yet, hardware nobody documented. Send me the specifics and I'll tell you within a day whether it's tractable and what it would take.