Machine Learning Engineer
Fuzzy Labs · Feb 2024 – Mar 2026
- Co-led a RAG system over sensitive UK government legal documents.
- Used Qdrant as vectorDB and moved retrieval from semantic-only to hybrid (BM25 + dense) with reranking, fixing exact statute/clause lookups cosine search missed; measured the gain on NDCG@10 and MRR against a 660-query eval set built with government lawyers.
- Served it with vLLM behind an API-gateway auth layer that rejected unauthenticated queries before any LLM spend, added semantic caching to serve repeated queries without re-hitting the LLM, with Langfuse tracing, cutting query latency ~16%.
- Fine-tuned a pre-trained LLM (Mistral) on domain-specific UK legal data using QLORA (4-bit) to adapt it efficiently for training and production serving; evaluated the fine-tuned model against the RAG baseline and recommended against shipping after inference cost and latency did not justify the marginal business gain.
- For a real-estate internal-document chat product, designed and shipped a graph RAG system (PostgreSQL with pgvector) to replace underperforming hybrid RAG after diagnosing the data as inherently relational.
- Owned the evaluation, built the eval set, implemented hard-negative mining, and wrote LLM-as-judge scoring on context and answer relevance, run through the team's eval-gated CI before release.
- Managed conversation context with query rewriting, HyDE, and reranked chunk shortlisting, plus a sliding-window memory that I extended with LLM summarization once the window alone lost earlier context.
- Explored inference-optimization techniques, quantized an LLM to GGUF (llama.cpp) for on-device mobile deployment in a POC, evaluating the cost and privacy trade-offs of moving generation to the edge.
- For an e-bike rental platform, built an event-driven pipeline with Kafka to stream and process a steady flow of rental, return, and availability events in near real time, with services communicating over gRPC and REST, Dockerised and deployed on AWS behind a load balancer, with blue-green deployment and Prometheus/Grafana monitoring.

