Ilya Trapeznikov

• · Split the system in two: an astronomy core on native Swiss Ephemeris (C) computes the exact chart, and the AI (Gemini) only receives finished numbers and writes the human reading, so it never makes the stars up.
• · Isolated the native C ephemeris add-on onto Node functions, working around its incompatibility with Vercel's Edge runtime.
• · Built a single surface (web and a Telegram Mini App): content unlocks instantly over Supabase Realtime, with three ways to pay (Telegram Stars, crypto, bonus balance).
• · Built a data-driven landing page catalog for organic search: ~2,700 unique pages across 10 languages, 10,000+ URLs.

Built with: Next.js 16 · React 19 · Supabase (Realtime) · Gemini · swisseph (C) · React Three Fiber · grammy

• · Architected a pipeline that turns hundreds of open Telegram channels in over a dozen countries into one board: deduplication by normalised text, a cheap prefilter, rule-based parsing into strict JSON, merging of the same object across channels and stripping of personal contacts. The live loop runs without a key to any external model.
• · Split the work between rules and a model: price, period and currency are parsed deterministically, while verticals are annotated by Claude Code subagents in batches. Two independent passes agreed on 98.6% of 286 shared posts, and full annotation of the rejects measured the candidate net's real recall at 78%.
• · Set a release gate for parsing fixes: every change runs read-only against the whole corpus, with a random control of about 8,000 posts that must not change at all.
• · Engineered the listing fingerprint from region, district, bedrooms, price and area: cards mixing two countries dropped to zero, and the largest merged group went from 973 posts to 239.

Built with: Python · httpx · BeautifulSoup · Telethon (MTProto) · SQLite · Supabase · Vercel (static PWA) · Claude Code subagents

• · Isolated each company's data at the PostgreSQL RLS level: in a multi-tenant system, one company's data stays separate from every other's.
• · Built a step-by-step builder for 44-FZ procurement specs with hierarchical document sets and export of ready-to-sign files in Word and Excel.
• · Deployed an nginx gateway on a Russian VPS so Vercel's cloud can reach state registries past geo-blocks.
• · The AI fills values strictly inside a fixed template, each with its justification, and leaves anything it cannot support blank for a person to check.

Built with: Next.js 16 · React 19 · Supabase (Postgres + RLS) · Gemini · docx · exceljs · Vercel

• · Built a matrix backtest engine on vectorized computation: thousands of parameter combinations in one pass, about 77 times faster than the obvious loop (roughly 13 minutes down to 10 seconds per instrument).
• · Built an overfitting-rejection pipeline: Walk-Forward, Out-of-Sample, Monte Carlo and two more independent checks against luck.
• · Modeled leveraged trading through virtual capital with margin rebase, which the underlying library does not support, with no look-ahead.

Built with: Python · Streamlit · VectorBT · pandas · NumPy · ccxt · SQLite · Pine Script

• · Designed a no-code scene builder so staff assemble and change the screens themselves, with no developer and no code.
• · Built a master-and-slave architecture: one server keeps physically separate screens in step through WebSocket rooms, and it all runs offline.
• · Implemented state recovery: after a power cut a "blind" screen comes back on its own to exactly what it was showing.
• · Already live in a museum, running the "Siege of Leningrad" exhibition.

Built with: Vue 3 · Node.js / Express · Socket.IO · PostgreSQL 15 · Electron · Docker Compose

• · Designed the group-planning model: a "worst-case visa" calculation across the whole group's passports and destination ranking by a composite metric.
• · Built a server-side LLM layer with function-calling and streamed answers from an AI guide grounded in the user's own data; all paid service keys stay on the server, never in the browser.
• · Tightened the access policies (RLS) so users can never see each other's data.

Built with: Flutter Web · Supabase (Realtime) · Vercel Functions · Gemini · Telegram Bot API

• · Stitched CRM deals together with a closed procurement-and-warehouse system into one BI layer by reverse-engineering an undocumented API.
• · Fixed a systemic error of naively summing quantities that pushed contract fulfillment past 100%.
• · Restored ~90% of deals that arrived with no stage or program by reconciling against internal ids: 14 of 16 contracts saved that would otherwise have dropped silently from the report.
• · Sped up the build with parallel requests (ThreadPoolExecutor) and overnight data preparation.

Built with: Python · Streamlit · pandas · requests · SQLite · nginx

Ran the full project cycle: client, proposals, specs, delivery, on-site installation. Automated the manual parts for myself: the commercial-proposal generation, and a layer over the CRM that removed manual deal tracking (the Deal Intelligence dashboard above).

Built tools and algorithms for analyzing financial instruments and trading at different risk levels (the basis for TRPZLabs Quant Radar).

Helped launch the company from zero and built its infrastructure, still running today. Led the technical staff, scoped projects and tender specs, and handled suppliers, clients and public authorities.

Maintained a fleet of about 200 workstations, ran user support (service desk), deployed and supported OS and software, and helped roll out new internal services.

Hands-on PC repair and retail sales (top salesperson, 2007).