Code experiments

Small prototypes,honest writeups.

I build small open prototypes when an idea will not leave me alone. I keep them deliberately narrow, ship a working version in a weekend or two, and then write up what I actually learned. Not what worked on the marketing site, what worked in the code.

None of these are products. A few have grown into client work, most stay as labs. The point is to learn out loud and leave a trail other engineers can follow.

12
Labs in flight
Experiments, not products yet
36
Technologies touched
Mix of new and familiar
5
Categories explored
AI, real-time, edge, more
4
Past prototype
Beta or stable today

Featured experiments

~/labs/ai-code-review
beta
AI/ML

AI Code Review Assistant

A small prototype that turns a GitHub pull request into a contextual code review. I wanted to see how far you can push an LLM with a tight feedback loop, structured diffs, and a real linter running alongside. The result is a tool that catches the boring stuff (naming, dead branches, missing null checks) and surfaces the interesting stuff (design choices, hidden coupling) before a human even opens the PR. It is not a replacement for review, it is a sharper first pass.

Read writeup
~/labs/realtime-collaboration
alpha
Real-time

Real-time Collaborative Editor

An exploration of CRDTs and operational transforms for building Google Docs-style collaborative editing experiences. I wanted to understand the offline story end to end, not just the happy multiplayer path. This prototype runs on Y.js with a tiny WebSocket relay, persists everything to IndexedDB on the client, and reconciles edits when you come back online. The interesting work is in the seams: presence, undo, and how you tell a user that their last 12 minutes of typing just collided with someone else's.

Read writeup
~/labs/edge-ml-inference
experimental
Edge Computing

Edge ML Inference

I wanted to see how close to zero you can push ML inference latency by running small models at the edge instead of calling a cloud API. This prototype runs quantized ONNX and TensorFlow.js models on Vercel Edge and Cloudflare Workers, with WebAssembly handling the heavy lifting. The headline result is sub-10ms classification globally with no cold starts. The interesting result is everything that does not work: anything above ~50MB, anything that needs a GPU, anything stateful.

Read writeup

Filter by category

Filter by status

01.lab
AI/MLbeta

AI Code Review Assistant

A small prototype that turns a GitHub pull request into a contextual code review. I wanted to see how far you can push an LLM with a tight feedback loop, structured diffs, and a real linter running alongside. The result is a tool that catches the boring stuff (naming, dead branches, missing null checks) and surfaces the interesting stuff (design choices, hidden coupling) before a human even opens the PR. It is not a replacement for review, it is a sharper first pass.

OpenAI GPT-4TypeScriptGitHub APINext.js
Explore lab
02.lab
Real-timealpha

Real-time Collaborative Editor

An exploration of CRDTs and operational transforms for building Google Docs-style collaborative editing experiences. I wanted to understand the offline story end to end, not just the happy multiplayer path. This prototype runs on Y.js with a tiny WebSocket relay, persists everything to IndexedDB on the client, and reconciles edits when you come back online. The interesting work is in the seams: presence, undo, and how you tell a user that their last 12 minutes of typing just collided with someone else's.

Y.jsWebSocketReactIndexedDB
Explore lab
03.lab
Edge Computingexperimental

Edge ML Inference

I wanted to see how close to zero you can push ML inference latency by running small models at the edge instead of calling a cloud API. This prototype runs quantized ONNX and TensorFlow.js models on Vercel Edge and Cloudflare Workers, with WebAssembly handling the heavy lifting. The headline result is sub-10ms classification globally with no cold starts. The interesting result is everything that does not work: anything above ~50MB, anything that needs a GPU, anything stateful.

ONNX RuntimeVercel EdgeTensorFlow.jsWebAssembly
Explore lab
04.lab
Developer Toolsbeta

Visual Query Builder

A drag-and-drop interface for building complex database queries without writing SQL by hand. It introspects Postgres metadata, lets you wire joins visually, and emits clean SQL or Prisma queries on the other side. I built it because every internal tool I have shipped eventually grows a 'just let me filter this differently' request, and at some point a UI is cheaper than another ad-hoc endpoint. It is not a replacement for SQL, it is a way to let non-engineers ask 90% of the questions without involving an engineer.

Reactdnd-kitPostgreSQLPrisma
Explore lab
05.lab
AI/MLalpha

AI Form Generator

You describe a form in plain English, the tool generates a fully validated React component with proper field types, accessibility, and a Zod schema for validation. I wanted to test how reliable structured LLM output has become in 2025-2026 for a tightly scoped generation task. The answer is: very reliable, as long as you constrain the output schema aggressively and validate the model's response before rendering it.

OpenAIReact Hook FormZodshadcn/ui
Explore lab
06.lab
Data Visualizationbeta

Streaming Analytics Dashboard

A real-time analytics dashboard pulling events through Kafka, aggregating in ClickHouse, and pushing updates over WebSockets at sub-second refresh rates. I built it to learn how the columnar-store world has changed: ClickHouse is dramatically faster than the OLAP setups I was running five years ago, and it makes a single-node setup feel like a small data warehouse. The dashboard itself is intentionally simple, the interesting part is the pipeline behind it.

Apache KafkaClickHouseReactD3.js+1
Explore lab
07.lab
AI/MLexperimental

Voice-First Interface

An experiment in voice-driven web UIs with real-time transcription and natural language commands. The trigger was watching how often I reach for keyboard shortcuts in tools I use daily, and wondering whether voice could be a faster path for some of those interactions. The answer is: sometimes, on the right device, in the right room. This prototype combines Whisper transcription with an intent classifier and a small command router. It is interesting, not a product.

Web Speech APIWhisperOpenAIReact
Explore lab
08.lab
Developer Toolsalpha

Zero-Bundle Components

A small component library that ships zero JavaScript to the client by leaning on React Server Components and modern CSS. I wanted to test the hypothesis that most of what we put into client bundles in 2026 is theatre. The answer is: a lot of it is. Tabs, accordions, dialogs, dropdowns, even sliders can run with no JS at all if you accept some constraints. The constraints are real, but they are smaller than I expected.

React Server ComponentsCSSNext.jsTypeScript
Explore lab
09.lab
Real-timeexperimental

Local-First Sync Engine

A sync engine for local-first applications. The client owns a SQLite database in the browser (via WASM), every write goes there first, and a small server reconciles edits across clients. I built this to understand whether the local-first promise (offline, instant, multiplayer, but still synced) is actually achievable for everyday CRUD apps in 2026. Short answer: yes, but conflict resolution is still where the design work lives.

SQLiteWASMIndexedDBWebSocket
Explore lab
10.lab
AI/MLexperimental

AI Testing Agent

An AI agent that writes and runs end-to-end tests by reading your app like a user would. It opens pages with Playwright, identifies elements visually and structurally, drafts test cases, and runs them. The honest result: it is great for happy paths and visual regression, mediocre for complex flows, and useless for anything that requires domain knowledge it does not have. Worth it as a first sweep, not as your test suite.

OpenAIPlaywrightTypeScriptComputer Vision
Explore lab
11.lab
Edge Computingbeta

Edge Rate Limiter

A globally distributed rate limiter built on Cloudflare Workers and Durable Objects. Token-bucket and sliding-window algorithms, sub-5ms decisions almost anywhere in the world, per-IP and per-key configuration. I built it because most rate limiters I have seen in production are either too slow (round-trip to a central Redis) or too lenient (eventually-consistent counters). Durable Objects let you do consistent counting at the edge, which is the actually interesting bit.

Cloudflare WorkersDurable ObjectsTypeScriptHono
Explore lab
12.lab
Data Visualizationalpha

Embedding Explorer

An interactive 3D visualiser for vector embeddings. Paste a corpus, pick an embedding model, watch UMAP project it into 3D, and click clusters to inspect what is inside. I built it as a debugging tool for RAG systems, because the fastest way to understand why a retrieval is bad is to look at the embeddings themselves. It turns into a teaching tool too: showing someone the cluster of 'shipping' tickets sitting next to 'returns' explains a vector database faster than a slide deck.

Three.jsUMAPOpenAI EmbeddingsReact
Explore lab

Keep reading

From experiment to production

Where these prototypes lead next.

Shipped work

The projects that started in a lab and grew into something real. Case studies of what made it out of the prototype phase.

See selected work

Writing and insights

Longer-form notes. Patterns, tradeoffs, and opinions earned from shipping these experiments and watching them break.

Read insights