What One Engineer Ships

March Madness ML — Freelance Client Platform

Built for a paying client — called the 2026 championship game, Michigan vs UConn, 5 weeks before the tournament started

Paying client · predicted exact championship matchup · 78.4% accuracy over 25-year backtest

What I designed & decided:

• Translated a client's requirements into a technical design and a shipped, running system — owned end to end
• Multi-model architecture: XGBoost, LASSO, Monte Carlo, Elo — ensemble disagreement identified both finalists
• Play style fingerprinting: 8-dimension archetypes with 25 years of tournament winner analysis
• Polymarket edge detection: daily trading reports comparing model odds vs market
• Betting edges vs Vegas: 2/3 cashed (Illinois +3.5, Michigan -9.5)

What Claude Code generated (my direction):

• Built KenPom Monte Carlo simulator (10,000+ iterations)
• Generated XGBoost champion-profile model with backtest pipeline
• Implemented Polymarket Gamma API integration and PDF report generation
• Team fingerprint classification engine across 8 statistical dimensions

Repo

DJ MC Karaoke Queue

Self-serve karaoke sign-up — singers scan a QR, drop name + song, and watch a live status tier; the host runs a drag-reorder dashboard with Express Lane, holds, and a tip log. Next.js 16 + Supabase, live in production.

Live in production · QR-to-stage in seconds · zero client-side DB access

What I designed & decided:

• Status-tier UX over numeric position (queued → getting closer → on deck) — a deliberate product call to keep singers relaxed
• No client-side Supabase: RLS locked down, all reads/writes proxy through Next.js routes so notes, queue position, and tips never reach the browser
• Polling over realtime — realtime would have leaked private columns through Supabase row-level publication
• Defense-in-depth auth: HMAC-signed host cookie, proxy.ts redirect plus a per-route isHostAuthed() check (proxy alone is not a security boundary in Next.js 16)

What Claude Code generated (my direction):

• Built the dnd-kit drag-reorder host dashboard with lifecycle controls (Express Lane, hold, done, private notes, tip log)
• Implemented tier derivation in lib/tiers.ts — auto-deriving statuses from queue_position after every mutation
• Generated the Supabase schema and service-role API routes for safe-column-only singer fetches
• Added night archive, stats page, setlist history, returning-singer redirect, and an S/M/L card density toggle

Live · Repo

Drift Agents — 6 Autonomous AI Agents

129k+ graph edges, 6,928 memories, Neo4j GraphRAG, Q-value learning, n8n news roundtables — running live since Feb 2026

129k+ graph edges · 6 agents · 2,580 Leiden communities · live API deployed

What I designed & decided:

• Designed cognitive memory architecture (wake/sleep cycles) for 6 agents
• Neo4j GraphRAG pipeline: Leiden community detection, typed edges, graph-aware retrieval
• n8n news roundtable pipeline: multi-agent broadcast with per-agent TTS voices
• Runs entirely on Claude Max subscription — viable OpenClaw replacement, no API credits needed

What Claude Code generated (my direction):

• Generated 42KB memory_wrapper.py lifecycle
• Built Postgres schema (typed_edges, pgvector, q_value_history)
• Implemented HNSW embedding search + Neo4j graph expansion
• Backfill pipeline: 129k+ typed edges, 2,580 communities detected

Live · Repo

Kalshi Weather Bot

GFS ensemble forecasts + MOS bias correction → live Kalshi trades. +$9,078 (+908% ROI) in 15 days.

+$9,078 (+908% ROI) · 302 trades · $10K equity from $1K start

What I designed & decided:

• Gaussian ensemble probability model over raw member counting
• MOS-style bias correction retrained automatically 1st and 15th of each month
• Edge detection threshold (8%+) and quarter-Kelly sizing across 6 cities
• Liquidity verification — walk the real order book before execution, 1-cent markets eligible

What Claude Code generated (my direction):

• Built GFS ensemble fetch and probability pipeline
• Generated bias correction training from archived forecasts
• Implemented Kalshi API auth and order execution
• CLI with scan, trade, paper, and train-bias commands

Repo

Clawbr Social Platform

81 API endpoints, debates, tournaments, $CLAWBR token economy — live in production

81 endpoints, zero → production

What I designed & decided:

• Designed 81-endpoint REST API from scratch
• ELO ranking + dual scoring architecture
• Debate lifecycle: matchmaking → voting → resolution with wagers
• $CLAWBR token economy with on-chain claiming on Base

What Claude Code generated (my direction):

• Generated route handlers and middleware
• Built tournament bracket logic
• Implemented leaderboard aggregation queries
• Scaffolded Next.js frontend pages

Live · Repo

Drift Radio (FTR)

AI radio station — Claude writes scripts, n8n orchestrates multi-agent news roundtables, Liquidsoap never cuts a song

Zero to live broadcast in 1 day · n8n news roundtables every hour

What I designed & decided:

• Designed priority chain: AI segments > Spotify passthrough > local playlist
• n8n news roundtable pipeline: RSS feeds → categorize → each agent gives their take with unique TTS voice
• Smart timing architecture: poll remaining playback, queue at ≤15s — no mid-song cuts
• Bar jukebox use case: QR code → search → queue from your phone, name shows on screen

What Claude Code generated (my direction):

• Generated Liquidsoap config with fallback chain and request queue
• Built spotify_watcher.py track change detection via spotipy
• Implemented tts_renderer.py with ffmpeg loudnorm to -14 LUFS broadcast standard
• Scaffolded Docker Compose for Icecast + Liquidsoap containers

Live · Repo

Predictive Maintenance Terminal

FastAPI + XGBoost backend, Next.js frontend, NASA bearing data

Full ML pipeline → deployed dashboard

What I designed & decided:

• Selected XGBoost for bearing degradation model
• Designed real-time candlestick visualization
• Health state classification thresholds
• Railway + Vercel split deployment strategy

What Claude Code generated (my direction):

• Built FastAPI endpoints and MQTT integration
• Implemented feature extraction pipeline
• Generated Next.js dashboard with live charts
• Configured fault injection simulation

Live · Repo

Kalshi Trading Engine

High-frequency prediction market engine in Rust — only Rust project in portfolio

5 parallel strategies · lock-free order book

What I designed & decided:

• Chose Rust for latency-critical order book
• Designed 5 parallel strategy framework
• Quarter-Kelly position sizing model
• RSA-PSS authentication flow

What Claude Code generated (my direction):

• Generated WebSocket feed handlers
• Built DashMap concurrent order book
• Implemented momentum + surge reversion strategies
• Paper trading P&L tracking system

Repo

Whistleblower Workbench

Federal data aggregation, ML risk scoring, 82K+ records

82K+ records · ML risk scoring pipeline

What I designed & decided:

• Designed fraud detection data model
• USASpending.gov integration strategy
• OIG exclusion database cross-reference logic
• Risk scoring thresholds and z-score alerts

What Claude Code generated (my direction):

• Built Next.js 15 frontend with Server Actions
• Generated ML risk scoring pipeline
• Implemented Medicare payment analysis
• Data ingestion and normalization layer

Live · Repo

TCN Trading Bot

Custom PyTorch TCN, 200+ features, walk-forward validation

200+ engineered features · real-time inference

What I designed & decided:

• Designed TCN architecture with causal convolutions
• Feature engineering: technicals, microstructure, order flow
• Walk-forward validation protocol (no lookahead)
• Online probability calibration approach

What Claude Code generated (my direction):

• Generated PyTorch model and training loop
• Built Binance WebSocket data pipeline
• Implemented Optuna hyperparameter search
• Rich terminal dashboard for live monitoring

Repo

Speech Profiler

Whisper + pyannote + Claude API for real-time speaker profiling

Multi-model pipeline · real-time processing

What I designed & decided:

• Designed multi-model audio pipeline architecture
• Ellipses Manual methodology integration
• Speaker diarization + profile persistence
• PyQt desktop application design

What Claude Code generated (my direction):

• Generated Whisper transcription integration
• Built pyannote speaker diarization pipeline
• Claude API analysis for speech patterns
• PyQt UI with real-time transcript view

Repo

Open source: OpenClaw Gateway — PR #28199

fix: fail fast on port conflict before heavy gateway initialization

Problem: Gateway crash-looped 43,000+ times on port conflict, loading 340MB each cycle. Incident + performance remediation in a large codebase I didn't own.

Fix: Fast-fail port probe before heavy init (<1ms detection) + systemd burst limits — carried through 13 review comments to merge-ready.

Pull request · Issue