work

long-running task agent with a browser operator console. next.js + cf workers. doom-loop guards, mid-prompt interrupters, follow-up steering, real-time token tracking, replay-based forking, subagent canvas via task-tool subagents.

operator console + telegram bot + recheck cron for publicly exposed comfyui / ollama boxes. classic agent loop: scan → fingerprint → diff → notify. fastapi + sqlite + react/vite + apscheduler. URL-synced filters, saved searches, watch list, model catalog, history + diff detection.

agentic latex editor. live preview + agent that edits the latex tree on user instruction.

native windows speech-to-text app. push / tap right-alt to record → whisper / deepgram → paste. C# / .net + global hotkeys. ships releases.

comfyui custom node that preserves image quality on flux kontext edits — diffs the edit against the original and selectively merges only changed regions back, instead of degrading the untouched pixels through full re-encode.

patched router-for-me/CLIProxyAPI to translate cursor ide's claude / anthropic toolcall format into gemini's functionCall. fixed missing tool-name sanitization, added image-block support, full unit-test coverage.

offline retrieval-augmented generation with ollama + llamaindex, multiple data sources, streaming responses.

yolov8 detector + easyocr for indian license plates. end-to-end video → text pipeline.