Nick Tran

Independent · 2025–2026

Pine Peak Associates research platform

Research and governance platform for systematic trading strategies, with gated research, server-side agent authority controls and no LLM in the critical path. No capital-affecting or production action without human approval; bounded autonomy inside evidence gates for research and paper trading.

Python · Data platforms · AI agents

Pine Peak Associates was my independent engineering and quantitative research work, from 2025 until I wound it down in mid-2026 to join a wealth management startup.

Problem

I wanted AI agents to do the repetitive parts of strategy research (factor discovery, backtests, write-ups) without giving an agent, or a bug, a path to capital. That needed a research process that rejects weak signals on evidence, and authority limits that hold when the model ignores its instructions.

Work

Architecture

Pine Peak architecture A research layer feeds a deterministic orchestration layer, which feeds an execution layer. A human approval boundary sits between orchestration and execution. Research and paper trading run with bounded autonomy inside evidence gates. Research layer factor discovery walk-forward CV signal validation Orchestration deterministic no LLM in critical path gates, leases, audit log Execution paper live human approval boundary bounded autonomy inside evidence gates

Gate pipeline

Gate pipeline Five stages in order: signal validation, walk-forward training, dual backtest, paper trading, risk review. A candidate advances only on the evidence artifact for its current stage. Signalvalidation Walk-forwardtraining Dualbacktest Papertrading Riskreview deflated Sharpe · multiple-testing correction · out-of-sample seal

Agent authority

Agent authority enforcement An agent sends a signed request. The server derives identity from the signing key, then applies an authority envelope, a work lease, a dry-run default and kill switches before any tool runs. Agent signed request Server identity from signing key per-agent authority envelope leased work claim dry-run default tiered kill switches audit log Tool

Result

The research and governance machinery worked as designed, end to end. Audits of my own gate evidence surfaced problems I had not anticipated, and the rules that came out of them are in When Green Checks Lie.

Ask me about

  • Why deflated Sharpe and an out-of-sample seal, and where each still leaks.
  • Enforcing agent authority on the server instead of in the prompt.
  • Why there is no LLM in the orchestration path.

← All work