Projects – Nolan Cacheux
A complete project portfolio spanning production systems, pilots, research prototypes, and hackathon builds. Each case study makes the project status, my ownership, architecture decisions, evidence, risks, and limitations explicit. 48-hour team prototype built for the Mistral AI Worldwide Hackathon after selection among 7,000+ applicants. I fine-tuned four Mistral models with QLoRA, built structured evaluations, and helped integrate them into an interactive political simulation. Deployed personal research prototype combining LangGraph orchestration, hybrid RAG over SEC filings, market data, and a conversational agent delivered through a messaging channel on Azure Container Apps. Designed to support analyst workflows with cited evidence, not replace human financial judgment. End-to-end personal MLOps project for AI-generated image detection with EfficientNet-B0 and Grad-CAM. I built the training, FastAPI serving, MLflow registry, DVC versioning, drift monitoring, infrastructure, CI/CD, and documented dataset limitations. Full-stack NLP prototype turning YouTube comments into sentiment, topic, and evidence-backed summaries. I implemented the BERT and BERTopic pipeline, local Ollama summaries, FastAPI backend, Next.js interface, and live SSE progress. Production Google Chat enterprise agent for supplier and finance workflows. I led delivery from client discovery through Workday Finance integration, A2A-ready interface design, evaluation, security, deployment, and adoption. Designed for a 10,000-user target population and validated with a 2,000 concurrent-user load test. Enterprise agent for maintenance, safety, and compliance workflows in Google Chat. I led product and technical delivery across source governance, governed RAG, cited answers, safe fallback, evaluation, observability, and cloud deployment. ROI remains an explicit hypothesis: 14,500+ hours/year and approximately €540k of productivity reallocation. Production ML pipeline replacing manual forecasts across eight sales KPIs and 64 sports categories, with model benchmarking, feature engineering, MLflow, CI/CD, and automated delivery. Interactive educational web application for exploring neural network architectures from perceptrons to transformers. I built the 3D renderer, architecture editor, mathematical explanations, and live TensorFlow code generation. Browser engineering prototype exploring board-state parsing, WebWorker-based Stockfish analysis, and explainable position evaluation with TypeScript and Manifest V3. Presented as a controlled analysis and learning project, not as a tool for circumventing fair-play rules. Production multi-tenant quiz platform built and operated for 100+ concurrent users. I designed the authoritative real-time game engine, Socket.IO event model, JWT security, observability, CI/CD, and Cloudflare-backed deployment. Experimental agentic trading system with xAI Grok analysis, Interactive Brokers connectivity, real-time monitoring, automated execution capability, and decision logs. The case study emphasizes architecture, risk controls, human oversight, and the limits of LLM-generated financial decisions. 
