Thomas Bateman
Full-stack engineer|AI agents, ML pipelines, trading systems
New to Contra
Thomas is building their profile!
ML Training Pipeline — 300 Models, 68 Features, Live Accuracy Tracking
End-to-end machine learning pipeline that trains, validates, and deploys 300 LightGBM models
across 25 symbols and 3 time horizons.
The 10-phase pipeline:
Data collection from Kraken OHLCV + orderbook
68-feature engineering (momentum, volatility, RSI, candle patterns, orderbook depth, BTC
cross-asset, macro regime)
Dead zone filtering — removes 30-40% of noisy bars where price barely moved
Groq LLM feature pruning — a 120B model ranks feature importance
3-seed ensemble training with LightGBM
Meta-stacking with out-of-fold predictions + orderbook features
Walk-forward backtesting (65-73% accuracy per symbol)
Automated deployment via rsync to production
The dead zone innovation is key: by excluding bars where |return| < threshold (BTC ±0.31%,
mid-cap ±0.36%, small-cap ±0.41%), the model only trains on meaningful price moves. This is why
live accuracy holds at 62-64% instead of degrading to coin-flip.
Stack: Python, LightGBM, NumPy, Pandas, scikit-learn, Groq API
0
3
Full-stack SaaS architecture handling real money, real users, and real-time data in production.
What's under the hood:
• JWT authentication with email verification and password reset
• Tiered subscription system (Free/Starter/Pro/Elite)
• Stripe and PayPal payment processing with webhook handlers
• Real-time WebSocket streaming for live data
• ML model inference serving 25 markets simultaneously
• Admin dashboard with user management, signal quality metrics, and system health
• 3Commas webhook integration for automated trading
The API serves a React SPA with Babel CDN transpilation — zero build step, instant deploys.
Behind Nginx with SSL on Oracle Cloud, running as a systemd service within a 2GB memory budget.
Stack: FastAPI, React 17, SQLite, Stripe, PayPal, JWT, WebSockets, Nginx, systemd, Oracle Cloud
1
16
An AI agent that doesn't just answer questions — it works. 188 callable tools, 47 operating
modes, dual-LLM architecture (120B brain for reasoning, 70B executor for tool calling).
It scrapes job platforms for leads, scores them, writes personalized proposals, delivers the
code, and invoices the client. Fully autonomous revenue pipeline.
Capabilities include:
• Full codebase auditing (12-phase enterprise security protocol)
• Self-healing and self-evolving code
• ML pipeline training and deployment
• Multi-agent swarm orchestration
• Legal defense document generation
• Real-time web GUI command center
95 Python modules. Zero external agent frameworks — built from raw API calls.
Stack: Python, Groq API, FastAPI, asyncio, LightGBM, Rich TUI, WebSocket GUI
0
5
Dominus Axis (Live Trading Engine)
Built from scratch and running in production. Real-time ML-powered trading signals across 25 cryptocurrency markets.
The engine combines 15 physics-based analytical pillars (Shannon entropy, Fokker-Planck, Lyapunov exponents) with LightGBM ensemble models trained on 68 engineered features. Every signal is validated by a 120B parameter LLM before delivery.
Results (302,512+ scored signals):
1h accuracy: 62.0% (+0.088% avg return)
4h accuracy: 64.2% (+0.509% avg return)
Cumulative 1h return: +486%
Stack: Python, FastAPI, React, LightGBM, WebSockets, Kraken API, Groq LLM, Oracle Cloud, Nginx, Let's Encrypt
0
6