Self-hostable RAG workspace where every answer cites the chunks it came from and the retrieval trace stays inspectable. Classifies each query and routes it to a dense, BM25, or hybrid retriever before generating, and stores 768-dimensional embeddings in PostgreSQL with pgvector under an HNSW index. Includes a statistical retrieval-evaluation framework with a leaderboard, and ships as a Docker Compose stack with Prometheus and Grafana wired in.

• TypeScript
• Next.js
• RAG
• pgvector
• Docker

Offline-first LLM benchmarking platform for evaluating local models across quality, performance, and resource constraints through standardized workloads. Built around Ollama with typed Pydantic contracts, reproducible JSONL benchmark artifacts, hardware profiling via psutil, and model-selection workflows — no cloud LLM dependencies.

• Python
• Ollama
• Pydantic
• LLM Benchmarking

F1 race-strategy platform that reconstructs pit-wall decisions: given lap, tyre compound, and gaps ahead and behind, it returns an ATTACK / PIT / DEFEND call with per-factor confidence. Combines 1,000-iteration Monte Carlo simulations, 800+ historical races from 2000–2024, XGBoost models with SHAP explanations, and simulated telemetry, plus a conversational AI race engineer backed by RAG over a ChromaDB store.

• Python
• FastAPI
• React
• XGBoost
• RAG

Redis-compatible in-memory server in Java 23, speaking RESP over the JDK's blocking socket API. Implements the core command set plus hashes, lists, and sets with strict value-type checking, key expiration, MULTI/EXEC transactions, an append-only file with on-demand compaction, and master/replica replication with partial resync. Ships as a Docker image and a Helm chart for Kubernetes.

• Java
• TCP/IP
• RESP
• Docker
• Helm