Projects

Problem

Flat vector search does not scale efficiently without sacrificing recall or cost.

This becomes a bottleneck for large-scale retrieval systems where latency, cost, and recall must be balanced simultaneously.

System Overview

Two-stage retrieval architecture:

• •Compressed routing layer built using MiniBatchKMeans super-clusters
• •Exact reranking over raw child vectors

Engineering Details

• •Python backend using FastAPI
• •Multi-tenant SaaS architecture
• •Snapshot-swap indexing for non-blocking ingestion
• •Redis + PostgreSQL + Supabase

Design Constraints

• •Optimized for recall parity, not approximate degradation
• •Memory-bound scaling beyond ~5–10M vectors
• •Deterministic, reproducible search behavior

Problem

Most trading strategies fail due to overfitting and lack of regime robustness.

System Overview

• •Self-evolving research platform using genetic algorithms
• •Continuous experimentation pipeline

How It Works

Started with 50 hand-designed seed strategies. The genetic evolution engine mutated parameters — entry/exit signals, price thresholds, technical indicators, and position sizing — across generations, producing 60,000+ unique variants. Each was validated through walk-forward analysis and Monte Carlo simulation before retention or discard. No strategy was retained without surviving out-of-sample testing.

Engineering Details

• •Python backend with FastAPI
• •Background workers for evolution, monitoring, and paper trading
• •PostgreSQL (Supabase), Redis

Problem

AI coding agents (Claude Code, Codex, Cursor, Copilot) operate with broad file system and git access but no standardized audit trail or policy enforcement layer.

System Overview

Vendor-neutral control plane that intercepts and audits agent actions at the commit/PR layer, with a webhook-based GitHub event pipeline and governance dashboard.

Engineering Details

• •FastAPI backend
• •Webhook pipeline
• •CLI tool published on PyPI (v0.2.1)
• •Validated against real-world repositories