Python, FastAPI, MCP, SQLite, ChromaDB, LangGraph, Pydantic v2

Built a substrate that versions an AI agent's cognitive state rather than its source: every conversation turn, decision and tool call persisted as an immutable, content-addressed Merkle DAG in SQLite plus a content-addressable blob store. Ships as a Click + Rich CLI, a FastAPI gateway, and an MCP server (stdio/SSE) so agents can commit, branch and restore their own reasoning. Published on PyPI as tm-ai (v3.8.19); 944 commits, 1,707 Python files, 147 test files.

CVC, Google Gemini 2.0 Flash, Python

Ran a 51-agent space-crisis simulation (1 Mission Controller, 5 Captains, 45 Specialists) where every agent's reasoning is persisted as a real CognitiveCommit and all 51 share one persistent hive-mind memory. 51/51 agents activated, 4/4 experiments passed, 5/5 integrity checks, 128 hive-mind commits, 172.5s total runtime on a 1.2 MB Merkle DAG.

Python, torchcrepe, RAVE, JUCE 8, Core ML, MPS

Built a voice-identity transformation engine (not pitch correction): a torchcrepe F0 detector at 100 fps driving RAVE timbre re-synthesis through IRCAM pre-trained models, with provenance sidecars. Measured on Mac Mini M4 (arm64): CREPE RTF 0.48, RAVE sol_ordinario_fast RTF 0.08 - 12x faster than realtime; Core ML export under 5% CPU per stereo channel. Ships as a JUCE 8 AUv3 and VST3 plugin for Logic Pro, Ableton, Reaper and FL Studio.

Python, Streamlit, Google Gemini AI, NetworkX

Built hierarchical/graph planner with memory and tool registry (HTTP, SQL, file, search); supports MCP-compatible JSON tool schemas and function-call I/O. Added task decomposition heuristics, reflection, and retries; improved task success on internal benchmarks by ~25–40% vs vanilla single-step prompting. Implemented run logs, traces, and cost/latency budgets per subtask; pluggable retrievers and safety rules.

React.js, Node.js, Express.js, MongoDB

Built a web application to manage office spaces and bookings, reducing manual allocation time by 40%. Designed a seamless UI, improving user interaction satisfaction scores, with a robust backend handling 1,000+ daily requests.

Python, Streamlit, Google Gemini AI

Engineered an AI teaching assistant, automating lesson plan generation and saving educators up to 30% of preparation time. Included an interactive Q&A system, supporting 95% accurate responses and enhancing educator productivity by 20%.

Python, Streamlit, GenAI, YT Transcript API, G-Cloud

Created a website using Google Gemini Pro for note card generation from YouTube videos. Achieved 100% accuracy in generating 10-12 insightful note cards per video, enhancing user learning efficiency by 25%.

Python, Streamlit, Google Generative AI, pandas

Produced an AI chatbot for real estate queries, improving user engagement by 30%. Integrated PDF and CSV processing for property data, reducing query response time by 40%.

Python, OpenCV, MediaPipe, Streamlit, Google Gemini AI, pyttsx3

Developed an AI-powered workout assistant for real-time exercise tracking using computer vision and pose estimation. Real-time rep counting and posture scoring; voice feedback; designed for low-latency edge inference.