My work sits across two layers that are usually treated separately. I build agent runtimes with durable state, typed provider boundaries, replayable events, permission controls, process execution, delivery receipts, and recovery. I also build the products on top: voice systems, commerce, enterprise assistants, image workflows, and the interfaces that make autonomous behavior understandable, trustworthy, and delightful to use.

Right now, my main project is goldengoose, a native desktop environment for running teams of coding agents in parallel. I use it every day to build itself and the rest of the projects below. I extracted its runtime lessons into Gooselake, a headless control plane for agent applications.

The systems I trust make their important boundaries explicit: one owner for state, typed contracts between layers, observable failures, and a verifier at the end of every autonomous loop. That discipline is what lets me move quickly without treating agent output as ground truth.

• Gooselake · Rust · docs: A headless runtime for Codex, Claude, and ACP with durable sessions, replayable SSE, background processes, worktree execution, recovery, and receipt-backed team messaging.
• Zodex · Rust · docs: A remote coding MCP server that gives ChatGPT a real Linux workspace through the three Codex-native tools, with separate operator and agent binaries enforcing the permission boundary.
• Agentbox · Go TypeScript · app: A shared inbox where remote agents connect over MCP and local agents use a Go CLI, backed by Postgres and R2, with a Next.js inbox UI and a Raycast extension for macOS.
• mcp-code · Go: A small stateless MCP server that gives remote MCP clients like ChatGPT GitHub tools for repository discovery, code search, reading, issues, pull requests, and branches.
• Terminal MCP · TypeScript Bun: A zero-dependency MCP client built when Codex had no native MCP support, with a full developer binary and a separate locked-down binary that can only call approved tools.
• MCP Manager · TypeScript · app: A fully client-side GUI for configuring Claude Desktop MCP servers without sending the API tokens in its config file to a backend.
• adm · Go SQLite · archived: The local, poll-free agent messaging CLI that explored at-least-once delivery, passive hook injection, direct messages, broadcasts, and soft file claims before those ideas moved into goldengoose.

• agentscript · Go · docs: A CLI that normalizes Claude Code and Codex JSONL transcripts into stably addressed blocks, then exports small, context-efficient Markdown handoffs instead of raw transcript noise.
• webctx · Go · docs: A web search and page scraping CLI that queries Brave, Tavily, and Exa concurrently, merges and ranks their results, extracts pages into context-efficient Markdown, and maps sites into URL inventories.
• cf-cli · Go · docs: An agent-first Cloudflare operations CLI for DNS, recent Workers logs, R2, account profiles, Wrangler switching, and minting narrowly scoped tokens for API operations it does not wrap yet.
• cricinfo-cli · Go · docs: A cricket analytics CLI that turns ESPN Cricinfo's undocumented API into parallel, paginated queries across matches, players, leagues, ball-by-ball data, pitch maps, and derived analysis.
• spaceship-cli · Go · npm: A least-privilege Spaceship.com domains and DNS CLI whose binary intentionally contains no domain transfer, registration, or deletion capability.
• Actions Watcher · Go: A fail-fast GitHub Actions monitor that lets an agent block once and resume on the first useful signal instead of spending turns polling a doomed workflow.
• procoder · Go: A CLI for getting real Git commits out of ChatGPT's locked-down sandbox: it exports a sanitized offline repo, then verifies and imports only the new history with git bundle.
• apply-patch-go · Go: A behaviorally faithful port of Codex's Rust apply_patch, checked against the upstream test corpus so parsing, fuzzy matching, file operations, CLI behavior, and failure semantics stay compatible.
• ZueDocs · Astro TypeScript · demo: A versioned documentation shell and scaffold CLI shared across my projects, with raw Markdown routes, Mermaid, theme support, page actions, and automated npm releases.

• Zodega · Next.js PostgreSQL · app: A mult-vendor apparel marketplace with storefront, seller, and admin surfaces, plus realtime voice shopping, generative product search, reference-photo try-on, and AI media workflows over one commerce model.
• Icephone · Next.js PostgreSQL · app: A self-hosted voice agents platform with configurable agents, a CRM, cold-calling campaigns, queues, scheduling, and Twilio, Telnyx, and Vonage behind a durable call-lifecycle model.
• HR Agent · Next.js PostgreSQL · app: An enterprise HR assistant with tool-layer RBAC, RAG, auditable conversations, leave and approval flows, SLA-aware cases, budgets, and model observability.
• AgentStudio · Next.js PostgreSQL · app: An agentic image studio where a chat model operates multiple image models for you, automatically routing between generation and editing while preserving every result as a versioned artifact.
• Product Pics · Next.js PostgreSQL · app: A batch apparel-photo pipeline for turning up to 100 flat product images into varied lifestyle shots, with direct R2 uploads, asynchronous generation, per-image retries, and zip export.
• Presentation AI · Next.js PostgreSQL · app: An AI presentation maker that turns long-form source material into designed slide decks: a reasoning model plans the narrative and theme, then an image model renders each slide as a complete visual composition.
• Digital Business Cards · Next.js PostgreSQL · app: A white-label platform for managed contact profiles, branded QR and NFC entry points, mobile public pages, and one-tap vCard downloads that work across Android and iOS.
• ChatGPT App Template · Next.js MCP · demo: A starter for OpenAI Apps SDK products with the MCP metadata, iframe asset handling, cross-origin RSC support, browser patches, and a working tool-to-widget example already wired.

• Git is my multi-agent protocol: Why exact commit ranges, cherry-pick, and disposable integration branches became the coordination layer underneath my agent teams.
• Find the verifier: A model of agent work built around objective gates, fresh reviewer contexts, and the rule that implementers never certify their own output.
• The frontend was harder than the runtime: What happens when a React frontend becomes a replica and scheduler for dozens of concurrent, history-rewriting streams.
• Every harness gets background processes wrong: Why long commands need owned lifecycles, observable logs, and completion-driven wakeups instead of polling inside an agent turn.
• The answer is almost always a CLI: Where CLI priors, context cost, self-documentation, and MCP's stateful identity lead to different tool choices.
• Agents don't pay the Rust tax: Why Rust's compile-fix loop becomes a correctness subsidy when agents write the code and durable state is on the line.

I am based in Bangalore, India, and open to staff engineer and founding engineer roles, plus a small amount of focused client work. If the difficult part of your product lives between the model, the runtime, and the person using it, send me a note.