Panos

A linter for ML training runs — deterministic PASS/WARN/FAIL verdicts on datasets, tokenizers, and first-epoch logs. Catch doomed runs before they burn GPU time.

Automated failure-mode QA for text-to-speech systems — structural audio checks, equivalence-aware WER/CER, ASR-uncertainty quarantine. Method published with DOI.

MCP-enabled evaluation and claim-grounding toolkit for long-document RAG systems - retrieval evaluation (5 methods), claim verification, a local FastAPI app, a local stdio MCP server for Claude Code/Codex, plus the hosted BookProof demo.

Daily AI content-automation pipeline: n8n schedules a decoupled FastAPI service running a LangGraph state machine (validate→retry) with LangChain + Gemini multimodal captioning and Pydantic-structured output. Self-hosted in Docker.

eal-time Voice-Activity Detection (Rust). Audio front-end for AI companions, offloading speech processing from the main game engine.

Local Windows desktop app + MCP bridge that lets Claude Code/Codex safely assist with Gmail - read, cleanup plans, drafts, and RFC 8058 one-click unsubscribe - every destructive action is human-gated in the GUI. v0.1 developer preview, source-available.

Multilingual emotional speech datasets for TTS training

Production-ready open-source Rust server for Amazon Polly TTS integration. Built for both local and cloud deployment. Zero compilation errors, no warnings—this binary is a diamond. Stable, clean, and used in real environments.