Twenty years building and leading engineering teams. Below is the work itself — live products you can open, source you can read.
My focus is AI on both sides: embedding it into products (RAG, agentic systems), and using it to change how engineering teams build software. I host and speak on getting agents past the demo and into production.
Open to engineering leadership and Forward Deployed Engineer (FDE) roles. Reach me on LinkedIn.
- Languages — Python · Go · C · Java · TypeScript · Swift
- AI — Multi-agent systems · RAG and hybrid search · MCP servers · AWS Bedrock · Local inference (Whisper · CLIP · Ollama)
- Systems — Metal GPU compute · ARM NEON SIMD · Lock-free concurrency and virtual threads
- Web and mobile — Vue · React · Spring Boot · FastAPI · iOS (Swift, ARKit) · Chrome extension · Google Docs add-on
- Cloud and delivery — AWS Lambda · CloudFront · Netlify Functions · Docker images on GHCR · PyPI · App Store
Every product below was built solo, end-to-end — the same AI-native practices I bring to a team, applied at n=1. Several ship in Spanish or Japanese rather than in translation; I'm trilingual (native English and Spanish, advanced Japanese, JLPT N2).
lens — multi-agent AI audit engine, founded and built solo, live and taking paying customers.
- 200+ checks across 26 dimensions of a codebase and a live site (security, performance, reliability, supply chain, cloud cost, accessibility, SEO) — one reasoning agent per dimension, grounded in deterministic scanners, every finding cross-verified by multiple frontier models to cut false positives
- Also ships as an MCP server, so the audit runs inline in Cursor, Claude Code, or Windsurf before the PR opens
- One audit produces two reports (plain-English for founders, engineer-grade for the team), a full dependency and license inventory (SBOM, in CycloneDX or SPDX format), and a prioritized fix-it roadmap
Live: lens.supersmall.ai · sample report
Kunkun — Japanese grammar-checking SaaS, built and run solo: Grammarly for Japanese.
- Context-aware corrections — politeness level, JLPT focus (N5–N1), and learner-vs-native mode shape every response, and each fix carries a plain-language explanation, not just the fix
- One backend, three surfaces — website, Chrome extension, Google Docs/Slides add-on — all reading live subscription state from the same API
- Go Lambda on AWS Bedrock (Claude/Nova) for inference, Firebase auth, Stripe subscriptions
Live: kunkun.io
Panorama — turns an iPhone into a spherical-panorama camera, a modern rebuild of the Photosynth experience.
- Guided ARKit sweep auto-captures frames and corrects for the phone's own tracking error
- Stitches into a seamless 360° image on-device with a custom Metal GPU stitcher
- Fully local: no cloud, no accounts, no backend, nothing collected
App Store: Panorama: Spherical Camera
Gaijin Smash — bilingual (EN/JA) direct-to-consumer streetwear brand: real Japanese slogans with proper cultural context, not Google Translate.
- AI content pipeline (image generation, model photography, copy) with human review as the quality gate before anything goes live
- Built solo end-to-end: storefront, checkout, CloudFront-backed asset pipeline, admin dashboard
Live: gaijin-smash.net
All four are proprietary — source not public.
Local-first tools you can run and audit yourself.
| Project | What it is | Stack |
|---|---|---|
| corpus | Ask natural-language questions over your own notes, PDFs, and docs. Hybrid semantic + keyword search with auto-tuned fusion, multi-hop reference expansion, a retrieval eval harness with a CI gate, and a 7-tool MCP server for Claude Code. Storage, index, search, and re-ranking are local; embeddings call your chosen provider. pip install corpus-rag |
Python RAG MCP |
| pdfcracker | Recover passwords from your own encrypted PDFs on macOS. All encryption revisions (R2 to R6), 15+ attack modes, GPU acceleration via Metal, ARM NEON SIMD, distributed cracking. Zero dependencies. | C Metal SIMD |
| imageclust | Clusters photos by what they're about, not just how they look. CLIP ViT-L/14 embeddings, Ward hierarchical clustering, and Ollama-generated labels. Entirely on-device. | Go React CLIP |
| cleancut | Drop in a video, get back a cleaned .mp4: profanity muted, explicit scenes cut. Layered local AI stack (Whisper, Ollama, NudeNet + LLaVA, HF audio models). Every cut is auditable. |
Python Whisper Vision |
| rehearsal | Practice spoken answers and get AI feedback, entirely on your machine: Whisper transcription, signal analysis of pace, pauses, fillers, and prosody, then a local LLM scores content and STAR structure. A Japanese mode adds a STAMP-aligned proficiency estimate. | Python FastAPI Ollama |
| claude-code-skills | 12 drop-in Claude Code skills for dev workflow and AWS ops. Auto-discover your AWS resources at runtime, no config files. | Tooling AWS DX |
| TPSGenerator | Java load tester for HTTP APIs on a Java 21 virtual-thread engine: stable, ramp-up, spike, and custom traffic patterns, chained scenarios, lock-free HdrHistogram metrics, circuit breaker, real-time resource monitoring. Pairs with TPSGenerator-Server. | Java Concurrency |
| online-storefront | Vue 3 + TypeScript storefront for small selling events: order form, Stripe Checkout, and a signed webhook that emails the customer and logs the order to a Google Sheet. No database, no server; fork it, edit one config file, deploy to Netlify. | TypeScript Vue Stripe |
- ELC: Production AI Agents, Beyond the Demo — hosted, July 2026. Where agents actually run in production versus stay stuck in pilot, and how you monitor a system with no single correct output.
- SuperSmall: Vibe-Code to Production — featured speaker, June 2026. Taking AI-generated prototypes (Lovable, Cursor, Replit) to production: auth, databases, deployment, performance.
Find me on LinkedIn.