Founder and Engineer · 2025–Present
A paper-stage system with live-grade architecture
Paramo is a systematic trading research system that I design and direct, built with AI coding agents. This page covers the controls that keep its results honest, the technical problems solved along the way, and how it is built. It has not traded real capital and makes no performance claims.
How It Is Built
Paramo did not exist before AI coding tools made it possible for one person to build it. The platform spans data ingestion, storage architecture, bias correction, pattern recognition, and a shared core for backtesting and paper trading. I architect the system and direct AI coding agents to write the code, and their output is verified by tests and CI before it counts. The architecture decisions, research logic, contextual judgement and quality control are mine. AI is a force multiplier here, not a shortcut, and the guardrails below exist because agent-written code needs them.
Controls and Key Technical Challenges
Controls Enforced in CI
Look-ahead scanning, a sizing-order check and a golden-regression baseline are enforced in CI, so the guardrails do not depend on anyone remembering to run them. The AI coding agents work behind ask-first gates and a pre-push audit hook.
Lookahead Bias
Backtest results were implausibly strong. Diagnosed forward-looking price information leaking into entry logic, enforced strict temporal data boundaries, and re-ran the full backtest suite so results reflect only information available at entry time.
Multi-Source Integration
Four financial data APIs with different schemas, coverage gaps, and undocumented inconsistencies. Directed the build of each ingestion pipeline and cross-checked data across sources before any analysis ran.
Shared-Core Architecture
The backtest engine and the paper-trading engine share the same core logic but have different runtime requirements. Designed a modular shared core where all strategy logic, pattern scoring, and exit management runs identically in both modes, with thin mode-specific adapters handling only the data layer.
Computational Scalability
Running full pattern analysis across 5,000 symbols with a 30-month lookback was projected to exceed 30 hours of runtime. Resolved by redesigning the screening pipeline to apply hard gap and price filters, dramatically reducing the symbol pool by 90% before hitting the expensive computation layer.