Ilia Mechkarov

• Data Engineer · HedgeServ

  November 2022 - November 2025
  • Developed ETL pipelines
  • Automated accounting processes using Python and libraries like NumPy and pandas
  • Led a project that migrated on-premise data and workflows to AWS - using step functions to automate data flows and S3 to store larger data files as we onboarded larger clients

• Data Advisory Intern · Deloitte

  May 2022 - August 2022
  • Built automatic indexing of an SQL Server database with Python, constructing a keyword catalogue out of plain text mainly using pandas and NumPy

• Capitol Tickers

  2025

  A multi-source ETL pipeline that tracks every public company and industry U.S. officials name in public - and what they personally own - modeled in Postgres for timeline analysis.

  • Python
  • ETL
  • Data Pipeline
  • Web Scraping
  • Postgres
  • Supabase
  • LLM
  • Typescript

  A data pipeline that ingests the public remarks and financial disclosures of U.S. officials, extracts the publicly traded companies and industries they mention, classifies the sentiment of each mention, and persists everything to a relational schema for timeline analysis. It pulls from several very different sources: Roll Call / Factbase transcripts (scraped from a paginated, server-rendered search API with requests + BeautifulSoup), Truth Social posts (pulled from the Mastodon JSON API, defeating Cloudflare's TLS fingerprinting with curl_cffi's Chrome impersonation), and OGE financial-disclosure PDFs (Form 278e holdings and 278-T transaction reports, parsed with pdfplumber). An LLM extraction stage turns the raw text into schema-conformant JSON - correcting speech-to-text errors, mapping each company to a real ticker, and tagging sentiment - which loaders then upsert into Supabase. The payoff is cross-source entity resolution: an official's disclosed stock holdings are linked by ticker to the same companies they praise or attack in public, all on one timeline.

  Highlights

  • Multi-source ingestion: paginated transcript API, Truth Social's Mastodon API, OGE disclosure PDFs, and raw transcripts - each normalized to a shared {event, segments} shape
  • Bypasses Cloudflare TLS fingerprinting with curl_cffi (Chrome impersonation) where plain requests/curl get 403'd
  • Normalized Postgres schema (events / entities / mentions / holdings) with unique constraints for idempotent re-loads and indexes on ticker, event, and entity
  • Incremental, stateful scraping (state files track the last run) so scheduled pulls only fetch new posts
  • Cross-source entity resolution links officials' disclosed holdings to the companies they name in public via shared ticker keys

  View Code →

• CS2 Mirage Entry-Route Analysis

  2026

  A data pipeline that maps opening-kill hotspots and T-side entry routes from pro CS2 demos.

  • Python
  • Spark
  • AWS S3
  • Data Analytics

  A data pipeline for Counter-Strike 2 demo files. A FACEIT scraper pulls a player's recent de_mirage matches via the FACEIT Data API, decompresses the demos, and a parser extracts every opening kill along with the attacker's approach path. The extracted data - opening-kill density, entry routes by side, and the hottest entry callouts - is written up as a short findings article on which spots and weapons win the opening duel on T-side Mirage.

  Highlights

  • Calls the FACEIT Data API directly (no brittle browser scraping) with retry/backoff
  • Rust-backed demoparser2 extracts opening kills and per-tick player paths
  • Canvas overlay with world-to-radar calibration, additive heatmap, and animated routes
  • Filter by player and side; hot-callout stats computed live in the browser

  Read the findings →

• Vim Muncher

  2026

  A Pac-Man-style arcade game played entirely with real Vim motions, with an A* solver that scores optimal shortest-motion paths between pellets.

  • TypeScript
  • Vite
  • Algorithms
  • Vim

  A browser arcade game where you navigate a maze using real Vim motions (w, b, %, ci[, and friends) instead of arrow keys. The interesting data work is in the game model itself: the maze is represented as a grid the engine reasons over, and an A* solver computes the optimal route between pellets using Manhattan distance, so the game can score how close a player's path is to the theoretical shortest sequence of motions. Game settings persist locally between sessions. A global per-time-control leaderboard is designed on a hardened Supabase schema - row-level security on every table and security-definer RPCs so scores can't be spoofed - but it is not wired up yet. Built with TypeScript + Vite and deployed on Netlify.

  Highlights

  • Maze modeled as a grid the engine reasons over for movement, collisions, and bracket-jump (%) targets
  • A* pathfinding with Manhattan distance scores optimal shortest-motion routes between pellets
  • Local persistence keeps player settings across sessions
  • Secure leaderboard schema designed (RLS + security-definer RPCs to prevent score-spoofing); not yet live

  View Code →

• Don't Say The Same As Me

  2026

  A realtime multiplayer party game where the database is the game engine - room, player, and answer state synced live across up to 7 players via Supabase Realtime.

  • TypeScript
  • Next.js
  • Supabase
  • Realtime
  • Postgres

  A browser party game for up to seven players: everyone answers the same prompt and scores only for answers nobody else gave. The interesting work is in the data layer - the entire game is a server-authoritative state machine living in Postgres. A rooms table drives the lobby -> question -> reveal -> finished lifecycle, with the round timer, current question, and a jsonb question plan all held server-side so no client can race ahead. Player answers are stored twice: the raw text exactly as typed, plus a normalized form used for matching, so near-identical responses collide correctly at reveal. Supabase Realtime publishes the rooms, players, and answers tables, so every client stays in sync without polling, and unique constraints on (room, round, player) enforce one answer per player even under concurrent submissions. Built on Next.js and deployed to Netlify.

  Highlights

  • Server-authoritative state machine in Postgres - round phase, timer start, and a jsonb question plan live in the rooms table, not the client
  • Dual-column answer storage (raw + normalized) so fuzzy-equal answers are detected as collisions at reveal
  • Live sync over Supabase Realtime publications across rooms / players / answers - no polling
  • Unique constraint on (room, round, player) guarantees one answer per player under concurrent writes

  View Code →

No projects match that filter.