Hobie One

A native desktop environment for working with an AI agent as a persistent partner rather than a per-session tool. I designed and built the entire system: a Rust/Tauri shell with a React/TypeScript front end, backed by a Node.js/TypeScript MCP service that exposes context tools to the agent. It maintains persistent cross-session agent memory with scope-routed handoffs, so a fresh session arrives already knowing the working state.

The interaction loop runs in real time: pointer-level deixis on a shared visual canvas, heartbeat and acknowledgment contracts, file-based event monitors, and an in-app browser that the agent can read and fill under human control. I use it daily as my primary working environment, which means every design decision has been tested against real use the same day it shipped.

A MCP server built solo under a timed hiring-challenge deadline: a multi-tenant member-data API exposed as agent tools. Every tool call passes through a tenancy gate, so one tenant's agent can never read another tenant's data. The build is spec-exact against the MCP standard, with a CLI demo client, Docker-verified builds that run clean in-container, and a suite of 18 tests, 7 of which are wire-level protocol tests covering the initialization handshake, tool listing, and malformed-call rejection.

The project went from an empty repo to and working, reviewed piece within the challenge window, and it is the piece I point to when someone asks whether I can ship production-ready agent infrastructure alone and fast.

A multi-tenant AI chatbot engine built for FwdEd. The back end serves many customer organizations, each with isolated data and its own assistant configuration.

I built the full stack, including the back-end API and an embeddable chat widget written in Preact within a Shadow DOM boundary. Hence, it drops into any customer page without style or script collisions, and a customer dashboard in Next.js 14 with pluggable AI components. Each tenant composes the assistant behavior they need rather than getting a single hard-coded bot.

The architecture keeps the AI layer swappable per tenant, which is what makes the engine a platform rather than a single chatbot. I designed and built the system end to end.

A research instrument that judges whether a written record structurally earns its own conclusion. Instead of fact-checking content, the engine reads the architecture of a claim: how evidentiary weight is distributed versus how prominently things are told, where support loops back on itself, and whether the verdict rests on a structure that can hold it. Its output is deliberately conservative, sustaining or withholding judgment, and its ceiling rule means it never accuses.

I built the method and the multi-agent evaluation harness around it, including blind scoring panels with pass bars committed before results, five-scorer rows with agreement statistics, and per-model validation, about 170 blind scorings to date across two LLM substrates.

In an early trial on a real decades-old criminal case, the engine withheld judgment on a conviction that was later overturned in the real world. The same harness discipline, parallel subagents with structured outputs and pre-committed bars, is what I bring to any agentic evaluation problem.

An AI admissions assistant for a real service business that turned website conversations into qualified leads.

A chat agent speaks with prospective students and gathers the information an admissions team needs, building a structured lead list from those conversations. A soft-call flow, then works that list and transfers interested prospects to an admissions rep.

My work was the conversational core. I took it over, hallucinating in live chats, with no retrieval behind it and a person hand-managing prompts to keep it on the rails. I sat with the real conversation flows first, then rebuilt it in FastAPI around a proper RAG pipeline with held conversation memory, so that heavy prompts stopped being resent when they did not need to be.

The rebuild cut hallucination rates by roughly 80% and reduced token spend, iterating against real conversations rather than a demo script. Customers do not care about the stack. They care that the answer is right.