Projects

Loading projects…

Computer Use Runtime

Model-free replay for software without an API

0 model calls during replay on a synthetic back-office UI

Details
Problem
Back-office software without an API forces a model to drive the screen on every run.
What I Built
Built four policy-gated discovery tools, a compiler to SHA-256-versioned capabilities, semantic targeting, and operator handoff in the same browser.
Impact
Against a synthetic back-office UI, replays with zero model calls, survives reversed table rows and columns, and passes 67 local tests; the fixture defines 12 scenarios, and a live discovery run with an injected page note still completed without an unregistered application call.
TypeScriptPlaywrightAI Agents
Code →

careeroc

Evidence-backed career workspace

118/130 public-sample regression cases passed, with 0 unsafe mutations

Details
Problem
Resume versions and recruiting history drift apart without evidence.
What I Built
Built a C# and PostgreSQL workspace over versioned resumes, recruiting history, and user corrections, with human review before any update.
Impact
The public sample passed 118/130 deterministic workflow regression cases (90.8%) with 0 unsafe automatic mutations.
C#PostgreSQLAI Agents
Sample code →Sample demo →Case Study →

Boring Agent Memory

Local recall layer for AI agents

0.990 Recall@1 on 120 synthetic queries

Details
Problem
Agents need local recall without replacing source files or crossing workspace boundaries.
What I Built
Built SQLite FTS5/BM25 retrieval with provenance, freshness checks, redaction, and workspace filters.
Impact
Across 120 synthetic queries: 0.990 Recall@1, 0.995 MRR, and 0 privacy leaks.
PythonSQLiteFTS5
Code →

LLM Serving Benchmark Harness

LLM serving benchmark that audits its own numbers

Identified a request-admission bottleneck in a local Ollama run

Details
Problem
Inference comparisons fail when workloads, hardware, or run artifacts differ.
What I Built
Built open-loop load tests, SLO goodput, KV-cache estimates, GPU telemetry, and claim and ranking audits.
Impact
Its published local run showed near-flat throughput while requests meeting the latency budget fell from 97% to 3%; goodput traced this to one-at-a-time request admission. A shared prompt prefix cut TTFT p50 6.6x through the server's prefix cache.
PythonLLM ServingEvaluation
Code →

SQL in C++

Paged storage engine with crash recovery

4 KB checksummed pages; recovery is tested at every traced crash point

Details
Problem
A database has to survive a crash at any write without losing or corrupting committed data.
What I Built
Built checksummed pages, persistent per-field B+ tree indexes, WAL-backed atomic single-writer statements, idempotent recovery, and copy-first migration from the v1 format.
Impact
Tests crash every traced write boundary, reject corrupted pages and WAL records, check B+ tree invariants, and run under AddressSanitizer and UndefinedBehaviorSanitizer in CI.
C++SQLStorage Engine
Code →

OAuth 2.0 Security Reference

Hardened OAuth sign-in for Next.js

8 threats in its threat model, each mapped to a control and a test

Details
Problem
Email matching can enable account takeover and leave OAuth replay gaps.
What I Built
Built provider-first identity, verified-email checks, one-time PKCE and state, rotated Redis sessions, and rate limits.
Impact
Tests cover callback replay, email-collision takeover, unverified provider email, concurrent callbacks, session fixation, malicious provider responses, brute force, and open redirects.
Next.jsPostgreSQLRedis
Code →

Other Projects

Agent Sandbox Eval

A Docker-sandboxed framework for reproducible tool-using agent evaluation with task manifests, JSONL trajectories, deterministic grading, replay, and failure analysis.

PythonDockerAI Agents
Code→

BookToday

A booking product for service businesses, in early access: a branded booking page and one workspace for services, staff, hours, and appointments, where guests book without a password.

Next.jsTypeScriptConvex
Demo→

Website Image Downloader

A bounded Next.js ingestion service with DNS pinning, redirect revalidation, SSRF and DNS-rebinding defenses, private-address blocking, timeouts, and deterministic security tests.

TypeScriptNext.jsSecurity
Code→Demo→

Tripemini

A 48-hour multimodal food and travel MVP with bounded Gemini inference, structured TypeScript outputs, validation, deadlines, rate limits, and concurrency controls.

TypeScriptNext.jsGemini
Code→Demo→

Sequential Recommendation Benchmark

A team benchmark comparing MostPopular, ItemCF, NeuMF, and SASRec for leave-one-out next-item prediction on the KuaiSAR short-video interaction dataset.

PythonPyTorchRecommendation Systems
Code→

cc-lite

An AlphaZero-style Xiangqi research baseline with a legal-move engine, PyTorch policy-value CNN, legal policy masking, PUCT MCTS, self-play, training, and engine distillation.

PythonPyTorchMCTS
Code→

Token Bar

A native macOS menu bar monitor for Claude Code and Codex with multi-account quota windows, capacity-weighted combined usage, API-equivalent cost estimates, and last-known values when a provider goes stale.

SwiftSwiftUIAppKit
Code→

Credit Card Optimizer

A browser-only rewards app that converts 27 cards' published rates into dollars back per $100, ranks them across 13 spending categories, and keeps the wallet on the device.

TypeScriptNext.jsZod
Code→

Mail Guard

A team IoT mailbox-monitoring prototype with ESP32-CAM firmware, a Next.js backend, image storage, and alert workflows.

Next.jsTypeScriptC++
Code→Demo→

BongoCat

A native macOS desktop companion that reacts to keyboard, pointer, mouse, and game-controller input using SwiftUI, AppKit, Metal, and Live2D Cubism.

SwiftAppKitMetal
Code→

SFML Graphing Calculator

A C++17 and SFML graphing calculator with expression parsing, RPN evaluation, interactive plotting, expression history, and headless tests.

C++SFMLParsing
Code→
Get in Touch