Activity Heatmap
- Jul 20, 2026 · 784.4M tokens
- Jul 21, 2026 · 2574.0M tokens
- Jul 22, 2026 · 149.4M tokens
- Jul 23, 2026 · 220.0M tokens
- Jul 24, 2026 · 1952.6M tokens
- Jul 25, 2026 · 315.6M tokens
- Jul 26, 2026 · no activity
@yauhenifutryn
Yauheni Futryn
Search hackers, navigate pages
Yauheni Futryn
Solo builder, mostly voice/audio AI and RAG pipelines. Recent stuff: a tool that turns a hypothesis into a fully-cited wet-lab experiment plan in 3 minutes (built in 24h, solo), and a Zotero citation-grounder that refuses to hallucinate references. Also placed #2 on a PII-reconstruction task at a European trustworthy-AI hackathon. On the client side I ship on-prem voice/RAG systems (vLLM, Dify, Qdrant) for companies that cannot send audio to a cloud API. I like writing down what is actually live vs still broken instead of demoing around it.
// null
// no data yet. connect GitHub to populate.
I build voice/audio AI systems, RAG pipelines, and the occasional research tool, usually solo, usually against a deadline. Most of what's below came out of hackathons or real client work rather than tutorials.
Sextant — an AI ops platform for contract research: give it a hypothesis, get back a fully-grounded experiment plan (protocol, materials, budget, timeline) with every claim cited to a real source. Four agents (Researcher, Skeptic, CRO Operator, Compliance) fan out and a fifth consolidates. Every correction a scientist makes turns into a typed rule that feeds the next plan. Built solo in 24 hours for Hack-Nation 5 (Fulcrum Science's "AI Scientist" track). Live demo.
CISPA Warsaw 2026 — competed in a European trustworthy-AI championship across three ML-security tasks: membership inference, PII reconstruction from an over-fit model, and LLM watermark detection. Placed #2 on PII reconstruction. The membership-inference entry came down to a parametric mixture model over per-sample loss distributions, and the single costliest bug was an image normalization mismatch that quietly wrecked the loss targets.
Community OS — a reusable event-intelligence pipeline for hackathon organizers: turns applications, attendance, and public GitHub/project evidence into cohort statistics and a partner-facing report, without reducing anyone to a single score. Built for START Warsaw. — 286 applicants, 83 accepted, 78 confirmed attendees.
Team zer0_day submission archive for a European trustworthy-AI championship: membership inference, PII reconstruction (#2), and watermark detection.
Grounds [Author, Year] placeholders in an LLM-drafted document against your Zotero library. Deterministic matcher; LLM only for optional disambiguation.
Reusable event-intelligence pipeline for hackathon organizers: applications, attendance, and project evidence into cohort stats and partner reports.
Air-gapped RAG stack (vLLM + Dify + Qdrant) built to run Qwen2.5-72B locally and scale from one A100 to a 4-GPU cluster with no code changes.
Production audio-intelligence pipeline for a leasing company: call recordings to a knowledge base to a live voice assistant over SIP telephony.
citeground — grounds [Author, Year] placeholders left behind by an LLM-drafted thesis or report against a real Zotero library, then writes native Zotero citation fields straight into the .docx. The matching core is deliberately non-LLM (fuzzy string matching against Better BibTeX), because hallucinated citations are exactly the failure mode this exists to prevent. An LLM only gets involved, opt-in, for the 5-15% of ambiguous cases.
On-prem voice and RAG systems — knowledge_agent is an air-gapped RAG stack (vLLM + Dify + Qdrant) built to run Qwen2.5-72B locally and scale from a single A100 to a 4-GPU cluster without code changes. leasing is a production audio-intelligence pipeline for a leasing company: raw call recordings in, a searchable knowledge base and a live voice assistant over SIP telephony out.
Solo, fast, and allergic to demos that quietly overstate what's real. Several of these repos ship a SHIP_REPORT.md or an explicit "here's what's live vs. still gated" note rather than pretending everything works. I'd rather say a feature is disabled pending a credential than let a reviewer find out the hard way.