Mustafa Mehmood
Applied Scientist
Designing, evaluating and fine-tuning agentic systems for high-stakes domains
Chat with Agent SherlockExperience
Applied Scientist
Thomson Reuters
- •Developed core agentic workflows for M&A due-diligence review in CoCounsel, TR's AI legal assistant, across hundreds of documents per deal, grounding agents via tools in expert-curated legal sources; achieves 90%+ recall of lawyer-flagged material risks.
- •Own evaluation methodology: defined grading criteria with lawyers and designed LLM-as-judge strategies calibrated to expert scores (0.78 Cohen's κ), detecting 87.6% of human-flagged errors; cut eval cycles from days to hours.
- •Built an orchestrator agent harness and skill library encoding lawyers' review methodology, plus MCP tools for citation verification, cross-reference resolution and redline comparison.
- •Compared context-engineering strategies for long documents (retrieval-then-reason, progressive summarization, sub-agent state passing), yielding up to 4× fewer orchestrator compactions without loss in recall.
- •Fine-tuned Qwen-32B for clause customization with QLoRA (SFT, then DPO on lawyer preference rankings), matching GPT-4o in lawyer review at 5×+ lower per-token cost.
LLM Training Data Specialist
Outlier AI
Data Scientist Intern
University of Nottingham
Education
University of Nottingham
Nottingham, UK
Key Modules
Projects
Selected work
Sherlock-AI: Character Emulation with LLMs
Can a fine-tuned LLM authentically emulate a literary character's personality, reasoning style, and speech patterns?
CourseCompanion: RAG-Powered Course Assistant
How can professors instantly create AI teaching assistants from their existing course materials?
DeepGaze: Autism Diagnosis via Webcam Eye-Tracking
Can affordable webcam-based eye-tracking replace expensive clinical equipment for early autism screening?
Publication
Peer-reviewed research
Pre-diagnosis for Autism Spectrum Disorder Using Eye-Tracking and Machine Learning Techniques
Mehmood, M., Amin, H. U. (2024)
This work presents a novel approach to early autism spectrum disorder screening using commodity webcam eye-tracking combined with machine learning. By converting gaze patterns into visual representations and applying both classical ML and deep learning classifiers, we achieved 92% detection accuracy — demonstrating that affordable, accessible tools can support clinical pre-diagnosis.View on Springer Nature (DOI)
Skills
Tap or hover a domain or project to filter
LLMs & Agents
Training & Evaluation
ML & Data
Infrastructure
About
I'm an Applied Scientist at Thomson Reuters, where I build agentic systems for M&A due diligence in CoCounsel, own their evaluation methodology, and fine-tune models for legal drafting. My work sits at the intersection of large language models and domains where accuracy isn't optional.
Before this, I created expert RLHF data at Outlier AI, writing multi-step tool-use trajectories and preference rankings for frontier AI labs. I also spent two years in computational neuroscience, building ML systems for neurological disorder diagnosis — work that led to a peer-reviewed Springer Nature publication.