I build RAG pipelines and agents that turn slow, manual document work into something that happens in seconds. Most of my focus goes to the parts that decide whether AI actually ships: retrieval, evaluation, and infrastructure that holds up on real data.
A selection of systems I've built across retrieval, agents, and the data infrastructure around them. Each one started as a real problem to solve.
Production system that processes complex financial contracts end-to-end using RAG and LangGraph agents on AWS Bedrock. Reduced a 30-45 minute manual review process to seconds at 95%+ accuracy.
Reads letters of credit (SWIFT MT700 advices) and their amendments with vision models, then checks bills of lading, invoices and certificates against the credit terms the way a bank examiner would under UCP 600. Deterministic code settles what arithmetic can decide; an AI examiner judges the page images, and every verdict is fingerprinted to the exact credit state and lands in a review matrix for human approval.
An MCP server over a DuckDB financial warehouse: exposes the schema as resources and guarded, read-only SQL tools so an LLM can safely answer analytical questions, with a per-table allow-list and audit logging.
Designing multi-step agent pipelines with branching logic, tool use, and human-in-the-loop review.
Building semantic search systems with embeddings, HNSW indices, and re-ranking pipelines that actually scale.
Catching behavioural regressions before they reach prod. Trace-level diffing, statistical tests, CI integration.
Production-grade Python services: async FastAPI, queues, caching, type safety end-to-end.
Pipelines that don't break at 3am. Schema design, batch & stream processing, knowledge graphs.
Deploying AI workloads on AWS with infrastructure-as-code, CI/CD, and full observability baked in.
Liverpool John Moores University · United Kingdom
School of Information Science · Rabat, Morocco