Co-founded. A local-first AI workspace built around one idea — *capture anything, anytime*. An agentic assistant that organizes your notes, photos, and links into a single window, with a research-and-memory brain working underneath.

Python / FastAPI backend, web frontend packaged as a macOS desktop app. Fuses a coding-agent cockpit for driving local projects with a multi-model brain — Deep Research, RAG memory, and an MCP capability layer — so one window steers both your code and any model.

Co-founded. A context infrastructure for AI-assisted engineering teams. Graph-based metadata store with code-anchored entities; an event-driven Context Bus ingests webhooks from Slack, GitHub, and CI; a Dynamic Rule Engine scopes team standards from org → file; MCP Server (JSON-RPC 2.0) endpoints serve structured context to AI tools.

Built Feature Rooms that auto-synthesize discussion threads into executable AI prompts, and a Living Documentation system that syncs branch-specific docs via tree-sitter AST tracking — with NLP extraction and confidence scoring for cross-branch semantic conflict detection before merge.

A framework for components that describe themselves to AI. Each Semant component renders the same UI for humans, then exposes its state and operations through four parallel channels — plain text, llms.txt, JSON-LD, and MCP tools — so agents can read and operate the page without vision models or DOM scraping.

Zero-dependency core with React and Vue adapters that share identical semantic output; Svelte adapter on the roadmap.

Co-founded. A full-stack SaaS that extracts "prompt playbooks" from developer interactions across Claude Code, Cursor, and Codex CLI, capturing only the developer input — never proprietary code.

Shipped multi-tenant with RBAC, custom scoring profiles, and team benchmarks (signal breakdown, adoption rates, trends, task-type distribution). Surfaces anonymized high-performing strategies as coaching material.

A local-first AI writing framework (CLI + Web UI) built for 500K+ word fiction. The data model is an 18-table SQLite schema (WAL, FTS5) with foreign-key entity linking across characters, locations, and plot arcs — so context assembly is deterministic.

Implemented a 5-layer priority-based context builder that walks the FK graph within a token budget, feeding LLM generation via YAML-templated prompts with SSE streaming and six generation modes.

A full-stack meeting planner with real-time venue discovery via the Google Maps API, multi-tier Redis caching, user auth, event management, and group voting. Tuned API performance through query optimization and caching strategy.