- Spreadsheet Agent
—
Exploring how far you can push AI agents in structured data environments. Sub-agent architecture (Explore, Plan, Verifier) with hybrid search (regex, semantic, numerical), semantic PDF parsing via DuckDB, and checkpoint-based undo. Built in Go with Claude API.
Go, Claude API, DuckDB
- Browser Agent
—
Fast, reliable web automation against adversarial UX. Replaces screenshot-based observation with pruned ARIA trees and structural page features, cutting per-action latency from 3-5s to 1.5s. Planner-executor architecture with a persistent skill library. 30/30 on adversarial challenge benchmark.
Python, Claude API
- Mise en Place
—
Hands-free cooking with AI voice agent. Step through recipes by voice, adapt portions, substitute ingredients from what's in your fridge, clarify steps mid-cook. Claude Sonnet voice agent with tool use controls the Flutter UI. n8n Track Prize, ElevenLabs Hackathon.
Flutter, Claude API, ElevenLabs
- Internal Developer Tooling
—
MCP servers (Hive, SelectDB, PostgreSQL) with encoded data access controls and schema context, giving analysts natural language access to production databases. Plus a Lark CLI (Go binary proxying to Lark API via Claude skills) for pulling data from operational sheets and docs.
Go, Python, MCP Protocol
- Harry Potter Portraits
—
Living portraits that see you and talk back. Ingest a YouTube video or photo, extract face + clone voice, generate persona via LLM, produce gaze-aware animations. TV displays a portrait that tracks the viewer via webcam and responds conversationally through a STT-LLM-TTS pipeline.
Python, fal.ai, LivePortrait, WebSocket
- Claude Radio
—
Voice-first multi-agent command center via Telegram. Attach to any running Claude Code session, send voice commands, get live audio updates. Chief of Staff LLM routes rambling voice input to the right agents. Spawn agents, approve diffs, fix CI - all by voice note.
Python, Claude API, Telegram