William Masih

AI Engineer intern at Amundi, Europe's largest asset manager. Working on agentic AI systems and context engineering for financial-domain tasks: multi-step LLM agents with tool use, retrieval, memory, and evaluation. End-of-studies master thesis internship.

• Designing agentic pipelines with tool-use and structured outputs
• Context engineering: prompt architecture, retrieval strategies, eval design
• Production-grade LLM systems with cost, latency and hallucination constraints

Won the AI hackathon organized by Sopra Steria, Crédit Agricole and AWS. Designed an anti-phishing solution for the banking sector combining multi-layer ML models with LLMs on AWS Bedrock.

• Multi-layer ML pipeline for phishing detection
• LLM layer on AWS Bedrock for contextual verification
• Banking-sector threat model, ship-ready UX
• Winning team against competing student teams

One-year internship at BMW's Autonomous Driving Campus in Munich. Worked across MLOps, data engineering, and generative AI for computer vision. Real production datasets, AWS infrastructure, close collaboration with the CV/perception team.

• Pipeline optimization: enhanced AWS data pipelines and optimized heavy SQL queries (materialized views) to process large-scale autonomous driving datasets
• Generative AI: implemented data augmentation with Stable Diffusion to improve the robustness of computer vision models
• Monitoring tools: developed Python Dash analysis tools and Grafana dashboards to monitor sensor data quality and validation workflows
• Contributed to active learning experiments with tracking models

One-month summer role. Real-time allocation of taxi drivers to SNCF requests under strict time constraints. Not technical — relevant as evidence of fast decision-making and composure under pressure.

• Real-time triage of urgent transport requests
• Composure and decision-making under high-pressure constraints

Repomind turns a repository into a versioned local memory snapshot, then extracts compact context packs for an AI agent. Instead of dumping a graph or the whole repo into a prompt, Repomind keeps rich memory locally and only sends the proof packet relevant to a specific question. Direct embodiment of context engineering as a discipline. ~10k LOC Python, full-stack: Tree-sitter parsers (Java/Hibernate), Oracle schema introspection, SQLite graph store, FastAPI backend, React frontend with 2D and 3D views, local heuristic chat over the indexed graph.

• Substrate snapshot model (graph + nodes + edges + inventories + manifest), versioned per commit_sha — not just a graph
• ask-snapshot CLI: routes a query to the right sub-context, returns a compact markdown/json proof packet ready for any LLM
• Evidence discipline: every answer separates observed / inferred / unknown
• Persona-aware answers (business / support / technical languages over the same evidence)
• Workspace model (system graph): code + DB + APIs + queues + auth + business glossary, not just files
• Targets enterprise stacks (Java/Hibernate/Oracle first) where naive code chat fails

Personal AI assistant I designed, built, and self-host on a Hetzner VPS. Multi-agent architecture with memory, RAG, scheduling, Discord interface, and a Flutter companion mobile app. Runs 24/7 as a systemd service. Hands-on ownership of a full agentic stack — from infra to agent orchestration.

• Multi-agent system (BaseAgent abstraction: Architect, Memory, Scheduler, Orchestrator)
• LLM orchestration via litellm across Groq, Gemini, MiniMax with fallback routing
• Memory stack: PostgreSQL + ChromaDB for semantic recall, custom GapTracker
• Mobile bridge (FastAPI) + Flutter companion app
• Self-hosted on Hetzner CX22: systemd, logrotate, nightly backups, fail2ban, ufw

Progressive NLP journey: a from-scratch Seq2Seq chatbot trained on Cornell Movie Dialogs, then BERT and GPT-2 fine-tuning, culminating in a functional RAG pipeline. Useful grounding for understanding what LLMs hide from you.

• Seq2Seq with attention trained on Cornell Movie Dialogs corpus
• BERT fine-tuning for intent classification
• GPT-2 fine-tuning for generation
• Functional RAG pipeline end-to-end

Small robot with real-time perception. Face verification using DeepFace (VGG-Face) on webcam frames + OpenCV color-shape detection (Haar cascades + HSV masks). Threaded verification loop with on-screen status (enemy spotted / sector clear).

• Face verification with DeepFace (VGG-Face)
• OpenCV Haar cascades + HSV masks for red/blue/green shapes
• Threaded perception loop
• On-screen status UI

Personalized image suggestion engine based on unsupervised clustering on a visual feature space. Explored different clustering algorithms and similarity metrics.

• Feature extraction on image dataset
• Clustering for suggestion grouping
• Similarity-based retrieval
• Evaluation of recommendation quality

Trading recommendation bot trained with AutoML on Ethereum market data and derived signals. Learned the hard way that most crypto ML papers are non-reproducible and that feature engineering dominates model choice.

• Custom feature engineering over ETH price + indicators
• AutoML model search and evaluation
• Backtesting framework

Experimental 2D RPG where the Game Master is an LLM. Turn-based combat with dice rolls, JSON-driven maps, LLM narrates within rule constraints. Explores narrative freedom bounded by deterministic game mechanics.

• Turn-based system with dice rolls & stats
• LLM-driven GM for narrative freedom
• JSON-based maps & interaction rules
• Unity + Python backend

E-commerce prototype for Pokémon cards built as three microservices (StoreAPI, CardAPI, UserAPI) with Docker Compose orchestration. Focus on clean architecture and scalability patterns.

• Three microservices: StoreAPI, CardAPI, UserAPI
• Dockerized with REST endpoints
• PostgreSQL + auth & sessions

Academic simulator for emergency-service routing. Multi-agent environment where RL agents learn dispatch strategies under time constraints. Evaluated against heuristic baselines.

• Multi-agent emergency-routing simulator
• RL agents for dispatch strategy
• Evaluation against heuristic baselines

Advanced data augmentation pipelines combining Segment Anything (SAM) for mask generation, ControlNet for conditional synthesis, and superpixel segmentation for region-aware augmentation. Related to my BMW generative-AI work.

• SAM for automatic mask generation
• ControlNet for conditional image synthesis
• Superpixel-based region augmentation

First version of my portfolio: 3D scene with a kung-fu figurine on a desk surrounded by CRT TVs, HL2-inspired UI overlays, camera transitions between sections. Replaced by v2 to showcase agentic AI and context engineering.

• React + Vite + R3F + GSAP
• Camera transitions between category nodes
• HL2-inspired UI windows with typewriter text

Completed the Deep Learning Specialization by Andrew Ng on Coursera / deeplearning.ai. Covers neural networks, hyperparameter tuning, regularization, optimization, CNNs, and sequence models.

• Neural networks and deep learning
• Hyperparameter tuning, regularization, optimization
• Convolutional neural networks
• Sequence models (RNN/LSTM/attention)

Master's at CPE Lyon, a French Grande École engineering school. Specialization in data mining, machine learning, deep learning, and reinforcement learning. Combines strong software engineering fundamentals with an AI/data track.

• Specialization: data mining, ML, deep learning, RL
• Core CS: algorithms, OOP, databases, distributed systems
• Math foundations: linear algebra, calculus, probability & statistics
• Diplôme d'ingénieur expected 2026

French classes préparatoires — intensive two-year math/physics/engineering foundation leading to the entrance exam for Grande Écoles. Known for rigour in math and scientific reasoning.

• Mathematics (linear algebra, analysis, topology)
• Physics and engineering sciences
• Problem-solving at engineering-school entry level

Specialization on applying ML techniques to quantitative trading. Covers time-series features, supervised learning for prediction, reinforcement learning for portfolio allocation, and financial data engineering.

• Financial time-series features and labels
• Supervised models for price prediction
• RL for allocation and execution
• Backtesting and metrics

Hands-on experience designing LLM agents with tool use, retrieval, memory, and multi-step reasoning. Context engineering: prompt architecture, retrieval strategies, eval design, failure-mode analysis. From Amundi, FREYA, and this portfolio itself.

End-to-end RAG: chunking strategies, embedding model selection, vector DBs (Chroma, pgvector), reranking, citation tracking, eval metrics (hit@k, MRR, nDCG). Built RAG for FREYA, this portfolio, and NLP chatbots project.

Deep learning across frameworks: TensorFlow, Keras, PyTorch. NN fundamentals, CNNs, sequence models, regularization, optimization (SGD/Adam), evaluation. Deep Learning Specialization (Andrew Ng) certified.

From OpenCV/Haar cascades to CNNs, DeepFace, Stable Diffusion, SAM, ControlNet. Strong at the data side: augmentation, active learning, dataset curation, drift detection (BMW).

Supervised & unsupervised learning, feature engineering, hyperparameter tuning, model evaluation. Still the right tool for most tabular problems. Used AutoML for trading bot, clustering for image recommendation, and classic ML for hackathon win.

SQL + Pandas for wrangling. Materialized views for query performance. Data-quality and validation pipelines. Dashboards in Grafana and Python/Dash (from BMW). Comfortable in AWS S3 + downstream processing.

Self-host FREYA + this portfolio 24/7 on Hetzner: systemd, nginx, Let's Encrypt, PostgreSQL, fail2ban, ufw, logrotate, cron backups. Comfortable operating my own stack without managed services. Docker for containerized services.

French native, English fluent (BMW internship 1 year in Munich, English-speaking team), Spanish professional, Japanese intermediate.