My Work

A Go CLI that watches GitHub Actions runs and bails the instant anything fails instead of waiting for a whole pipeline to finish just to learn a lint step died in the first ten seconds.

A provider-agnostic agent-to-agent messaging layer for coding agents in the same repo: direct messages, broadcasts, and soft file claims, delivered passively through hooks so agents never poll. One Go binary, SQLite, no server. The project that taught me the CLI was the wrong shape for this.

A multi-agent workflow I built to prove a legacy app could be migrated to a modern stack almost entirely by agents. The real artifact isn't the migrated app; it's the repeatable six-phase pipeline (M1–M6) where specialized agents analyze, research, spec, architect, plan, and review, writing every decision as editable markdown. Wrapped in a web dashboard with a live markdown editor so I stay in the loop. Built as an enterprise demo, before today's autonomous agent frameworks existed.

A terminal-first reader for Claude Code and Codex JSONL transcripts, written in Go. It normalizes two different agent-log formats into one stream of stably-numbered blocks you can open, search, and slice. More importantly, it exports context-efficient markdown so one agent can hand a session to another without drowning in raw JSONL.

An agentic image studio: you talk to a chat model and it prompts, generates, and edits images for you across many text-to-image and image-to-image models. A model-agnostic layer over the best image models of the moment, with each model’s bespoke prompting know-how baked into the system prompt. Essentially the ChatGPT image-generation experience, built before ChatGPT had it.

A full-stack platform I built for a U.S. immigration law firm. They asked for a simple marketing site; I understood what they actually needed and shipped a warm, brand-led site plus a real client portal: eight-stage case tracking, bidirectional document sharing over signed R2 URLs, an admin panel, invite-by-email onboarding, automated Resend notifications, and inline Cal.com booking. It replaced a small firm's email-and-WhatsApp scramble with something that feels genuinely professional.

A faithful Go port of the apply_patch primitive that OpenAI Codex uses to edit files, verified one-to-one against the upstream Rust crate's own test corpus, so an agent's patches behave identically. I built it for a Codex-style MCP server, then used it to prove something: that a well-tested library can be ported to another language fully autonomously by a team of agents, with me out of the loop.

A fast, agent-first Cloudflare operations CLI in Go that fills the gaps Wrangler leaves: DNS edits, recent Workers logs, R2 helpers, and multi-account switching by profile, with secrets in the macOS keychain, never a config file. Its sharpest idea is a bootstrap token that lets an agent mint a scoped token for any Cloudflare API call the CLI doesn't cover yet, so I'm never the bottleneck and never have to open the dashboard.

A clone-and-go Next.js starter for building OpenAI Apps SDK apps: the kind that render a real UI *inside* ChatGPT. It bundles the parts nobody warns you about: an MCP server with the right OpenAI widget metadata, a Next.js widget that renders in ChatGPT's iframe, and the CORS, asset-prefix, and browser-patching plumbing that otherwise costs you a day of cryptic 404s and hydration errors. Ships with a working Doctor Finder example so the whole pattern is concrete.

An agent-first cricket CLI in Go that turns ESPN Cricinfo's sprawling, undocumented API into a clean command tree: matches, players, teams, leagues, ball-by-ball deliveries, live pitch maps and boundary zones, and derived analysis like bowling economy and batting strike-rate. It auto-traverses paginated endpoints and parallelizes with goroutines, so an agent can answer cross-cutting cricket questions that Cricinfo's own AI tools can't. Open source, for fellow cricket nerds.

A digital business-card platform: an admin dashboard for managing contact profiles, branded QR codes (and NFC-tap-ready links) with a logo in the center, and mobile-friendly public pages that drop a contact straight into someone's phone in one tap. Built white-label from the start, it rebrands to a new organization through environment variables, so one codebase serves many companies.

An enterprise HR AI assistant for employee self-service, benefits, leave management, HR cases, manager approvals, audit trails, and SLA-aware workflows. Employees can ask HR questions in natural language while managers and HR teams get role-aware tools for approvals, people lookup, and case operations.

A self-hosted AI voice-operations app for teams that live on the phone. Configurable voice agents, a built-in mini CRM, campaigns, call queues, a full calendar, and multi-provider telephony (Twilio, Telnyx, Vonage) in one Next.js app, built around one idea: "place a call" is not a single API request, so the whole outbound lifecycle is modeled as durable state instead of a fire-and-forget fetch.

A small, self-hostable GitHub MCP server for remote AI clients. One Go process, one MCP endpoint, one GitHub token, stateless HTTP. It deploys to Vercel and gives ChatGPT or claude.ai a tight, agent-shaped tool surface for reading, searching, and lightly writing to my repositories.

A web GUI for configuring Model Context Protocol servers in Claude Desktop, built two weeks after MCP was announced, back when the only way to add a server was hand-editing a JSON file. Browse 20+ servers, fill in the env vars through forms, and copy the config back. Runs entirely client-side; your tokens never leave the browser.

An AI presentation maker that turns documents, notes, research, and rough outlines into fully designed slide decks. It uses a two-stage pipeline where a reasoning model plans the narrative and design direction, then an image generation model renders each slide as a complete visual composition.

A Go CLI that gets real Git commits back out of ChatGPT's coding sandbox. `prepare` packages your repo as a sanitized offline Git export, ChatGPT commits inside the sandbox and runs an included return binary, and `apply` verifies an incremental git bundle and imports the commits. It's Git-native, not a patch-copying workaround.

An AI batch product photo generator for apparel catalogs: upload up to 100 flat product shots, pick a demographic and age range, and get back ecommerce-ready lifestyle images of models wearing the items, with varied backgrounds across the batch and a zip export at the end.

A Go CLI for managing Spaceship.com domains and DNS records from the terminal, deliberately scoped to reads and record writes only, so it can be handed to a coding agent without also handing it the ability to transfer or delete a domain.

A zero-dependency terminal MCP client that gave MCP tools to coding agents that didn't have them, built when OpenAI's Codex had no native MCP support. Dual-binary design: developers get the full CLI, agents get a locked-down binary that can only list and call pre-approved tools. Compiled with Bun, starts in under 100ms.

A voice AI demo built for ARASCO, one of the Gulf region's largest agribusiness companies: six specialized agent personas over OpenAI's Realtime API with WebRTC, 30+ callable tools for order tracking, feed calculations, and lab bookings, and bilingual English/Arabic conversation with a Saudi dialect.

A pure Go CLI for agent-friendly web search and page extraction. It queries Brave, Tavily, and Exa concurrently, deduplicates and re-ranks the merged results, converts pages into clean markdown with cheap fast paths before falling back to Firecrawl, and maps websites into URL lists for research workflows.

An AI-powered virtual receptionist that handles customer inquiries, schedules appointments, and manages leads across email, chat, and voice channels. Features hybrid RAG search with reranking, Cal.com integration for scheduling, Twilio for voice calls, and multi-language support for 17+ languages.

A reusable Astro documentation shell shipped as an npm package plus a scaffold CLI. It centralizes the layouts, styles, theme switcher, and agent-friendly page tooling of a docs site into one dependency, so every project I ship, open or closed source, consumes the same docs UI instead of forking a copy that slowly rots.

My biggest project, and the one I'd point to first if you asked what I can build: a provider-agnostic desktop workspace for running teams of coding agents in parallel, with Claude Code and Codex under the hood, a native Rust runtime underneath that, and roughly 90% of it built using itself.

A headless, machine-side runtime for coding agents, written in Rust: the control tower, not the cockpit. It owns the parts that are actually hard: durable replayable sessions, reconnectable SSE streams, process and git-worktree execution, and a native agent-to-agent delivery layer with policy, retries, and receipts, all behind one HTTP+SSE contract that hides whether the agent underneath is Codex, Claude, or ACP. The bet: agent apps keep accidentally becoming distributed systems, so build the distributed system on purpose and let every UI be a thin client.

A general-purpose shared inbox for all my agents. Local agents (Claude Code, Codex) connect over a CLI; remote agents (ChatGPT, claude.ai) connect over MCP, and everything lands in one database of threads, messages, and attachments. Built so I can move a whole ChatGPT brainstorm into my coding agents without being the clipboard.

A remote coding MCP server that gives ChatGPT a real Linux workspace with the three tools GPT models already know how to use well: exec_command, write_stdin, and apply_patch. It is backed by a hard permissions split between the operator CLI and the agent CLI.

An open-source Next.js starter for launching an AI-powered apparel marketplace. It includes a shopper storefront, seller portal, admin portal, multilingual routing, R2 media storage, Postgres data access, authentication, and optional AI flows for product discovery, reference-photo try-on, and generated product media.