AI Engineer

A production algorithmic trading platform running four independent trading surfaces (prediction markets, FX, equities) with an LLM predict layer: Claude synthesizes RSS news and technical indicators (RSI, Bollinger, S/R, ATR) into narrative probabilities that drive live orders. Surface-isolated architecture — per-surface env files, bankrolls, kill switches, calibration curves — plus a subscriber distribution client with Kelly sizing and a multi-broker adapter pattern. Runs 24/7 on a $30/month infrastructure budget.

Accomplishments:
- designed and shipped a 10-gate env-var-driven risk chain (pair blacklists, direction filters, edge caps, cluster caps, daily-loss limits);
- built an audit-driven calibration layer — rebuilt the probability curve piecewise after Brier-score audits exposed 0.90+ band overconfidence;
- engineered shadow-outcome backfill resolvers populating would-have-been P&L on 640+ aborted signals ("saved by not trading" accounting per filter rule);
- operated the full stack solo: strategy, infra, payments, monitoring, incident response.

Tech stack: Python, FastAPI, Anthropic SDK (Claude), Kalshi/OANDA/IBKR APIs, PM2, systemd, DigitalOcean, JSONL logging

Overview: A fully autonomous daily intelligence product: a cron-driven engine gathers live odds data, Claude writes a quantified value-betting brief, and the pipeline renders and distributes it across email, Telegram, and a daily vertical video — with an auto-graded public track record so every output is falsifiable against reality. Zero human touch on a normal day.

Accomplishments:
- built the end-to-end content pipeline: data gather → LLM research → branded render → multi-channel delivery → automatic result grading;
- shipped the reel pipeline: LLM-designed HTML scenes → ElevenLabs TTS → karaoke captions → FFmpeg 1080×1920 → YouTube Data API v3 auto-publish;
- implemented deterministic de-vig fair-price math with the LLM constrained to narrative, not numbers;
- integrated paid billing (webhook-verified) and public results pages backed by an append-only ledger.

Tech stack: Python, Anthropic SDK, ElevenLabs, FFmpeg, YouTube Data API, Cloudflare Pages/KV, Resend, Telegram Bot API

Overview: A single pane of glass for a seven-product portfolio: every product emits HMAC-signed events into one Cloudflare D1 table via an ingest Worker; a token-gated read API and Pages dashboard render live revenue, funnels, and per-system health cards; per-surface kill switches route through a FastAPI control plane.

Accomplishments:
- designed the HMAC-signed, per-project-secret event ingest (server-derived project identity — a leaked emitter can't forge another product's events);
- built the D1 schema, read API with tiered bearer tokens, and PWA dashboard with offline shell;
- wired push-to-deploy CI across 7 private repos with idempotent process restarts;
- added Telegram fan-out so notify-worthy events ping the operator in real time.

Tech stack: Cloudflare Workers, D1, Pages, KV, TypeScript, Python, FastAPI, HMAC auth, GitHub Actions

Completed the University of British Columbia's Certified Ethical Hacker program (Sep 2024 – Jun 2025), covering the full EC-Council CEH domains: reconnaissance, network scanning and enumeration, vulnerability analysis, system exploitation, web application attacks, wireless and IoT/OT security, and defensive countermeasures — hands-on, lab-based offensive-security work. Applied directly in production: IDS hardening and network-security work on live industrial control networks in continuous-process manufacturing. Complements CompTIA Security+ and Network+ toward an OT/ICS security specialization.