Hi, I'm Sam

Computational linguist working on trustworthy AI in clinical and legal settings.Master's in Linguistics · University of Bergen
Clinical NLP · Ontology-grounded models · Trustworthy AI in healthcare
Open to PhD and research-engineering roles in Scandinavia

About

I focus on where AI systems fail in real-world use and how to make them reliable enough for decision-making.
My background is in linguistics, which shapes how I approach NLP. I start from how domain experts actually reason (clinicians working with ontologies, lawyers navigating case law) and build systems around that, rather than treating language as a purely data-driven problem.
Most of my work is in healthcare and legal contexts where the people using the tools need to trust what the model is doing.

Research Focus

My work is on making NLP systems reliable enough for expert decision-making in clinical and legal settings. Two problems recur across the projects on this page: whether a model's output can be trusted by the person acting on it, and whether the evidence for that trust survives evaluation choices that were not part of the original design.Current themes:• Clinical NLP
• Ontology-grounded language models
• Verification and confidence in LLM outputs
• Evaluation design for NLP systems
• Retrieval-Augmented Generation in regulated domains

Selected Research & Engineering Projects

Production Legal Document RetrievalRetrieval-augmented system for Norwegian legal documents, in production use at a legal-AI startup.Norwegian legal-AI startup, Innovation Norway-supported. Lawyers needed reliable retrieval across Norwegian case law and statutes, with a defensible audit trail.I worked with three lawyers and two engineers to define what "correct retrieval" meant on their material, then built the evaluation criteria out of that. Retrieval accuracy moved from 75% to 98% over several months of iteration. The system now processes 25 or more legal cases weekly in production.Architecture: Claude API, MongoDB, Azure. Prompts are version-controlled with rationales attached, so any output can be traced back to the retrieval decision that produced it. Most of the accuracy improvement came from tightening what "correct" meant, not from prompt changes.Links: GitHub (private)MedTermCheckVerification layer that checks LLM-extracted medical entities against ICD-10-CM and SNOMED-CT.Verification layer for LLM-extracted medical entities. LLMs will confidently output clinical codes that don't exist, or that don't match the source text.The system checks each extracted entity against ICD-10-CM (roughly 70,000 codes, local file) and SNOMED-CT (Snowstorm API), then scores confidence using four independent signals. If SNOMED is unavailable it degrades to ICD-10 plus source grounding: still useful, with fewer signals.Includes 20 annotated test cases with deliberate hallucination traps.Links: GitHub · Demo GDPR Article 9 Compliance CheckerDeterministic rule engine for auditing healthcare AI documentation against GDPR Article 9.Rule engine for healthcare AI documentation. Scans privacy policies and DPIAs against 42 Article 9 requirements, flags missing legal bases.Rules, not LLM classification. Compliance needs determinism and audit trails. "The system flagged this because rule 9.2.h matched" is defensible. "The model thought this was non-compliant" is not.YAML-based rules, evidence extraction, versioned decisions. English only. Limitations documented in the repo.Links: GitHub · Demo Medical Text Classification (Thesis)Neural classifier for systematic review automation, with a 2026 methodological extension.Neural classifier for systematic review automation. Small dataset (150 abstracts from a Cochrane dementia review), so I integrated NEO, a 1,611-concept neurological examination ontology derived from SNOMED-CT, into the feature space to compensate for the limited training data.Under 5-fold cross-validation, performance across 11 configurations clustered between 80% and 86%. Ontology enrichment changed which features the model relied on but did not reliably beat the bag-of-words baseline.In 2026 I extended the methodology to the Cohen et al. 2006 benchmark (2,744 abstracts across three topics). The finding sharpened: the effect of ontology-derived features on downstream classification is conditional on the evaluation design. A canonical bag-of-words gap on one topic holds under single-split evaluation and collapses under matched-corpus subsampling and 10-fold cross-validation at full sample size.Links: GitHub

Current Research

Research in progressEvaluation design conditions the expert-vs-auto MeSH gap: a con
trolled comparison of bag-of-words and BiomedBERT on the Cohen
benchmark
Working manuscript extending my thesis methodology to the Cochrane systematic-review benchmarks (Cohen et al., 2006). The effect of ontology-derived features on downstream classification is conditional on the evaluation design: a canonical bag-of-words gap on the Statins topic is robust under one evaluation protocol and collapses under matched-corpus subsampling and 10-fold cross-validation at full sample size.
Preprint forthcoming.

Background

Master's in Linguistics, Computational Track, University of Bergen (2024). Thesis on ontology-derived features for medical text classification. Teaching assistant for the Python and NLP courses during the Master's (91% completion rate).
Before Bergen: three years on the Volta River in Ghana, coordinating between Ghanaian operators and Chinese engineering teams on a underwater timber harvesting project. Currently teaching Python and NLP at the University of Ghana.

• View GitHub

About Me

Outside research, I spend most evenings at the piano working through jazz standards, in the kitchen trying to get a new recipe right, or reading science fiction. I've lived and worked in Ghana, China, and Norway, and I like projects where the technical work depends on collaboration across different backgrounds.

Contact

If you are working on clinical NLP or on verification and evaluation for LLMs in regulated settings, I would be glad to hear from you.[email protected] | LinkedIn | GitHub |Hugging FaceWorking on something in healthcare AI, NLP, or regulatory compliance? I'd be interested to hear about it.

My work is on making NLP systems reliable enough for expert decision-making in clinical and legal settings. Two problems recur across the projects on this page: whether a model's output can be trusted by the person acting on it, and whether the evidence for that trust survives evaluation choices that were not part of the original design.Current themes:• Clinical NLP
• Ontology-grounded language models
• Verification and confidence in LLM outputs
• Evaluation design for NLP systems
• Retrieval-Augmented Generation in regulated domains

Built with Carrd. Content © Samuel Okoe-Mensah 2026.