Amine El Idrissi — AI / ML engineer — LLMs, RAG, agentic systems

MSc Artificial Intelligence, University of Manchester (2026). I build production AI systems: LLM fine-tuning and RAG pipelines that shipped to 1,600+ users, multi-agent LLM systems with rigorous evaluation, plus computer-vision and knowledge-graph tooling. Looking for AI / ML engineer roles — available immediately, with full UK right to work and no sponsorship needed.

• Status: Available now · immediate start
• Right to work: UK · Graduate visa · no sponsorship needed
• Location: United Kingdom · Manchester
• Languages: English · French · Arabic

amyne.elidrissi@gmail.com · Download CV (PDF) · GitHub · LinkedIn

Experience

• AI Engineer Intern, FreshTrack (Jun 2025 — Aug 2025)
  • Fine-tuned an LLM to extract structured data from unstructured carrier documents (PDFs, emails), cutting manual data entry ~40% for 1,600+ logistics operators.
  • Built a RAG pipeline so operators can query carrier contracts, logistics rules and route history in plain English — lookups went from minutes to seconds.
  • Shipped an OpenAI-powered natural-language interface over live shipment dashboards, so non-technical staff get route status and carrier performance without SQL; reporting time down 35%.
• AI Engineer Intern, The University of Manchester (Jun 2024 — Aug 2024)
  • Developed LLAMA Gym CLI, an open-source tool for managing and comparing Llama 2 / Llama 3 models (7B, 8B, Instruct) to streamline LLM benchmarking.
  • Containerised and deployed it with Docker and Azure behind a custom API.
• Software Engineering Intern, Pratisoft (Jun 2023 — Aug 2023)
  • Worked in an agile team of 8 on a healthcare application, keeping it fully compliant with regulatory requirements and lifting user satisfaction by 25%.
  • Wrote 200+ tests and led debugging sessions, reducing software bugs by 30%.

Education

• M.Sc. Artificial Intelligence, The University of Manchester (2025 — 2026). Modules: Machine Learning, Deep Learning, Knowledge-Based AI, Data Engineering, Computer Vision. Dissertation: a controlled study of LLM trading-agent architectures (see projects).
• B.Sc. (Hons) Computer Science, The University of Manchester (2022 — 2025). Modules: Algorithms & Data Structures, NLP, Data Science, Software Engineering, Databases, Operating Systems.

Projects

• LLM Trading Agent Architectures (Agentic AI · MSc dissertation, 2026). Does an LLM trading agent's architecture matter more than the model inside it? A controlled backtesting study: three agent designs share one fixed model (Llama 4 Maverick), one system prompt, one information set and one execution engine, so every difference in behaviour comes from the design. A1 makes a single call per day; A2 adds a memory of recent trades and distilled lessons; A3 runs bull and bear analysts, a trader, and a risk manager that can only lower risk. Tested on TSLA and BTC from January to June 2026 with 3 seeds per cell: 18 runs, 2,727 logged decisions.
  • Risk fell monotonically from A1 to A3 on both assets: max drawdown down 36% on TSLA and 34% on BTC, turnover down 67% and 62%. The risk manager overrode 52% of the trader's proposals on TSLA (39% on BTC).
  • Returns improved along the same ladder (BTC −13.0% → −6.3%, while buy-and-hold lost 35%). A constant-mix control matched to each run's own exposure separates caution from skill: on BTC every agent beat its control, A3 by +6.6 pp; on TSLA none did. Bootstrap intervals include zero, so the ordering is reported as suggestive.
  • Leakage-safe and reproducible: the day-t information set is enforced in code and tested in CI, every LLM call is cached in SQLite, and all 2,727 decisions are committed with every prompt, response and risk override. The analysis reruns in about a minute without an API key.
  Result: Max drawdown cut 34–36%, turnover cut 62–67% (A1 → A3, both assets). Stack: Python, Llama 4 Maverick, pandas, NumPy, SQLite, pytest, GitHub Actions, Finnhub API. code, logs and results · full results report · watch the 13-second replay
• AI Receptionist (Bright Smile Dental) (Voice Agent, 2026). Voice receptionist for a UK dental practice: answers inbound calls, books appointments, handles FAQs and triages urgent cases end to end. Call it from your browser. Result: Live demo · call it from your browser. Stack: Next.js, Retell AI, GPT-5, TypeScript, Vercel. call the demo
• Mediterranean Culinary RAG (RAG Pipeline, 2026). Domain-specific question answering over Mediterranean culinary literature. Compares chunking, embedding, retrieval and prompting strategies in an end-to-end evaluation pipeline, so each design choice is backed by a measured retrieval and answer-quality score. Result: End-to-end retrieval + evaluation. Stack: Python, FAISS, BGE Embeddings, Gemini, Streamlit. open the demo
• Knowledge Graph Link Prediction (Graph ML, 2025). End-to-end pipeline benchmarking six knowledge-graph embedding models, packaged in Docker with GitHub Actions CI/CD and Optuna hyper-parameter search. Migrating R-GCN to sparse CUDA kernels halved training time and lifted ROC-AUC to 0.97. Result: ROC-AUC 0.97 · 2× faster training. Stack: Python, PyTorch Geometric, CUDA, Optuna, Docker, GitHub Actions. open the demo
• Visual Odometry for Autonomous Vehicles (Computer Vision, 2024). Monocular visual odometry on the KITTI dataset — estimates the vehicle trajectory from a single camera. Compares distance thresholding, nearest-neighbour and NNDR matching strategies. Result: Trajectory estimation on KITTI. Stack: Python, OpenCV, SIFT, KITTI. code
• Face Recognition & Completion (Classical ML, 2024). Face recognition and image completion on the ORL dataset using L2-regularised least squares — multi-class classification plus regression to reconstruct missing face halves. Result: Classification + regression on ORL. Stack: Python, NumPy, L2-RLS, Jupyter. code

Skills

• Languages: Python, TypeScript, JavaScript, Java, C++, SQL, Bash
• ML / DL: PyTorch, PyTorch Geometric, TensorFlow, scikit-learn, NumPy, Pandas, Optuna, CUDA
• LLMs & Agents: HuggingFace Transformers, Fine-tuning, PEFT / LoRA, Multi-agent systems, Tool use, Prompt Engineering, Evaluation & backtesting, OpenAI API, Gemini, Retell AI
• RAG & Retrieval: RAG, Embeddings, BGE, FAISS, pgvector, Chunking, Re-ranking
• NLP: BERT, Tokenizers, NLTK, Text classification, NER
• Computer Vision: OpenCV, CNNs, SIFT, Image preprocessing, Visual odometry
• Data & DBs: MySQL, PostgreSQL, MongoDB, Vector stores, ETL pipelines
• Infra & MLOps: Docker, GitHub Actions, Azure, CI/CD, pytest, Linux, FastAPI, MLflow
• Web: React, Next.js, Node.js, Django, Streamlit, Vercel
• Tools: Git, Jupyter, VS Code, Weights & Biases

This site works best with JavaScript enabled. Contact: amyne.elidrissi@gmail.com