My Work

Four production-minded systems spanning developer tooling, investment intelligence, private computer vision, and inspectable LLM orchestration.

Featured projects

RepoRipple↗

Sep 2026 — Present

Change-impact analysis for engineers and coding agents

A local-first change-impact analyzer that traces Git changes through Python, TypeScript aliases, and workspace packages, then recommends tests and review context for coding agents.

CLI, Action, MCP: 3 surfaces
deterministic core: Local-first
v0.4.0 release gate: 193 tests

Blast-radius tracing

Reverse-dependency paths connect each changed file to the code it can affect, then turn that evidence into focused verification guidance.

Agent-ready context

Typed, read-only MCP tools let coding agents forecast proposed edits or inspect a real worktree without receiving broad filesystem access.

Monorepo-aware analysis

Resolves TypeScript aliases and npm/Yarn workspace exports, while retaining dependency evidence for deleted and renamed source files.

Python
Git
GitHub Actions
MCP
OpenRouter
PyPI

Read case study →Install on PyPI ↗GitHub ↗v0.4.0 release ↗

SmartFolio↗

Apr 2026 — Present

AI portfolio and stock analysis platform

An investment-intelligence platform with deterministic portfolio analytics, position-level gains, performance history, and rebalancing previews.

analysis workflow: 8 stages
backend + frontend: 206 tests
pilot: 50+ users

Rebalancing previews

Exact-cent math compares target allocations and new-money plans before any change is applied to the portfolio.

Performance history

Dated valuations and cash flows drive estimated returns, drawdown, and benchmark comparisons with explicit data-coverage limits.

Position-level detail

Share quantities, cost basis, unrealized gains, and quote provenance stay connected without inventing missing data.

React
TypeScript
FastAPI
Vercel
Render
Neon Postgres
Plaid

Live site ↗GitHub ↗

What Is This↗

Jun 2026 — Present

Object recognition PWA and Chrome extension

A private visual identifier and Chrome guide: capture an object, confirm or correct its identity, and save verified results into searchable boards.

private recognition: On-device
public inference API: Hardened
two surfaces: PWA + MV3

Private recognition

MobileNetV2 runs on-device, while barcode and OCR recognition paths load only when needed on lower-end devices.

Hardened inference

A non-root FastAPI service enforces payload validation, rate limits, provider timeouts, bot protection, and structured logs.

Durable state

Versioned IndexedDB, legacy migration, row-level security, and Playwright accessibility and layout coverage protect data across phone and desktop.

Next.js
TypeScript
ONNX Runtime
FastAPI
Docker
PostgreSQL
Playwright

Live site ↗GitHub ↗

Token Optimizer↗

Jun 2026 — Present

Adaptive LLM workspace and Chrome extension

An inspectable AI workspace that chooses the leanest valid execution path, applies deterministic quality gates, and brings zero-call prompt preparation beside Gemini and ChatGPT.

adaptive execution: 3 paths
deterministic prep: 0 extra calls
two product surfaces: Web + MV3

Adaptive execution

Direct, compact-contract, and verified paths match workflow depth to the request instead of applying one expensive pipeline to everything.

Deterministic quality

Acceptance gates and conservative repairs validate explicit requirements without spending another provider call.

Inspectable by design

Streamed stages, usage details, safe public results, and a searchable code graph make each run and the system behind it easier to audit.

Node.js
Vercel
OpenAI
Groq
SSE
Zod
Manifest V3

Live site ↗Chrome Web Store ↗GitHub ↗Architecture report ↗

Engineering principles

Deterministic first

Code owns calculations and checks; models explain results and coordinate work.

Observable by default

Trace IDs, streamed progress, structured logs, and usage make each run inspectable.

Guardrails in the path

Validation, redaction, rate limits, fallback, and tests protect production behavior.