AI Engineer, Memon Systems Ltd
Jun 2026 – presentUK. Independent R&D lab. Remote
- Built a verifier that checks each sentence of an LLM answer as it is written and rewrites only the sentences that fail. It cut answers containing a wrong sentence by 86% when the model answered from its own knowledge, and by 48% with the source documents provided, refusing 7% and 4% of questions. Python package on PyPI. code and results
- Built a rewrite step for second-model checks on Claude and Gemini, so a rejected sentence is corrected against the source instead of deleted. It kept 71 of the 80 correct sentences the check had deleted, with wrong sentences held within the margin set before the run. results
- Tested Qwen 27B, self-hosted on one GPU, against the Jev and OpenAI decision APIs as a sentence judge, on 513 hand-graded sentences. On the first test it scored 93.8%, against 90.7% and 89.0%. method and results
- Built a rule-based router that sends each query to the exact legal source it cites before vector search runs, with no model call. 0 of 1,897 citations went to the wrong source, and 4.3% were refused. code and results
- Designed retrieval that keeps client data on the device and fetches public law through private information retrieval. It leaked 0 private tokens to the cloud, against 16 for standard cloud RAG. design and results