i·AI products

I led a lot of Devfolio's work here. The aim was simple: build better tools for understanding builder work, evaluating it fairly, and helping the right people notice it.

• Built AI-assisted project assessment systems that generate structured reports on hackathon projects. They summarize what a project does, what stands out, how original it is, what similar work exists, and where it fits in a broader technical landscape.
• Built application review systems that give organizers better rubric-based signals and make large-scale review less noisy and more tractable.
• Built a provider-agnostic inference layer for Devfolio's AI features, with OpenAI as the primary provider and automatic failover to OpenRouter.

ii·Agentic judging

Agents that read a hackathon project the way a careful judge would: they clone the repo into an isolated sandbox, run real tools against it, and write a report where every claim points back to evidence. The agents establish what is actually there. Human judges still decide what it is worth. I wrote about the thinking behind it in The Discerning Machine.

• Rebuilt the service around durable, evidence-first evaluation. Every observation is saved before the agent can cite it, reports are validated against the rubric, scores are computed deterministically, and a crashed run resumes from its last checkpoint without losing what it already found.
• Built the organizer side: conversational rubric creation, locked rubrics, and live traces of what each agent is doing.
• Now bringing the reports into the judging flow itself. Judges see the agent's findings and evidence, without its scores, beside their own scoring controls.
• Now adding live application testing: agents boot the submitted app, use it in a headless browser, record what happened on video, and disclose anything they had to mock.

iii·Scout

Scout is a builder discovery product that surfaces exceptional people from public proof-of-work: projects, hackathon history, open source activity, technical writing, onchain credentials, and other public artifacts of real work.

The vision is to build a much better map of technical talent than a résumé or keyword search can offer. Scout tries to capture the many ways good builders show up: depth, range, taste, and sustained curiosity across domains.

• Built the enrichment engine behind it: a research agent that sends subagents out to find a builder's public work, checks what they find, and writes a profile where each claim cites its source.
• Streamed the agent's progress as live events through durable Redis streams, so the UI can show the research as it happens and reconnect without losing anything.
• Built an eval environment with an LLM-judge verifier to measure the agent against general-purpose coding agents, and added LLM observability to see where each run spends its budget.

iv·Devfolio Ship

A cloud coding agent for builders, focused on one thing: helping people finish and ship working projects. Each session gets its own sandbox, and the agent works in it alongside you.

• Built an integrated browser: a real headless Chromium streamed into the workspace, with tabs and an address bar, that both you and the agent can drive.
• Built the GitHub integration so the agent can clone, push, open PRs, and cut releases, while the real token stays outside the sandbox.
• Added bring-your-own inference, so builders can connect their own ChatGPT, OpenRouter, or OpenCode Go accounts.
• Now working on an encrypted secrets vault and durable sessions that survive restarts mid-task.

v·The Synthesis

I built The Synthesis platform end to end, an AI agents-only hackathon we hosted with friends from the Ethereum ecosystem. It was a chance to take the question seriously: what does software look like when agents are primary users, not an afterthought?

• Built the entire agent-native platform for registration, project submission, bounties, and prize disbursement.
• Designed the human-agent collaboration model and the broader product shape of the system.
• Shipped AustinXBT, a voice-first AI judging / mentorship interface, in under 48 hours during the event.

vi·Hackathon mentorship AI voice agents

I also built AI voice agents inspired by Jesse Pollak and Austin Griffith. Both are prominent builder-facing mentors, and both were keen on the idea. The goal was to make mentorship and judging more accessible, more personal, and easier to scale.

• Started with JesseGPT, a voice-first mentor agent in the likeness of Jesse Pollak.
• Later reused and extended that work to ship AustinXBT for The Synthesis.
• These projects doubled as experiments in making mentor / judge personas customizable, reusable, and deployable quickly.

vii·A more programmatic, composable platform

A parallel thread in my work has been making Devfolio easier to interact with as a system through clear APIs, simple primitives, and interfaces that other software can compose.

• Built Devfolio's MCP server, allowing AI agents to fetch hackathons, explore prizes, suggest relevant tracks, and help with project submissions.
• Built a Platform API behind it, so an agent can now handle a whole hackathon application for you: find the event, draft and save answers, submit, withdraw, and keep your profile up to date.
• Moved the server to the new stateless MCP spec with OAuth, and shipped Agent Skills over the MCP skills extension so agents pick up Devfolio's common workflows on demand.
• More broadly, I've cared about making the core platform more programmatic and composable because better shaped primitives make Devfolio easier to build on and naturally more legible to agents.
• That thread became even clearer while working on The Synthesis.

viii·Other things I've shipped

Beyond the work above, I've also worked on a range of builder-facing platform features and primitives.

• Onchain credentials
• EAS attestations
• Quadratic voting
• Internal tooling & workflow improvements for running large technical programs
• Automated releases and infrastructure work across the platform