AI systems become useful when they produce evidence people can inspect, test, and trust.
My work spans AI model development and scientific applications, from implementing neural networks and language models from first principles to fine-tuning foundation models and developing biologically informed representations. I apply these methods to genomic interpretation, scientific evidence extraction, and secure clinical systems, with an emphasis on statistical foundations, reproducibility, and validation.
Scientific agents need structured sources, defined tasks, reviewable outputs, and clear links between input, reasoning step, evidence, and result.
AI-assisted clinical genetics requires secure deployment, controlled data flow, reproducible report generation, and human review by expert boards.
Systematic review, freedom-to-operate, and patent workflows need traceable extraction from article, claim, field, code, test, and human validation.
Agentic systems need credentialed access, clear user tasks, structured outputs, auditability, and product design that makes AI usable in real workflows.
PanelAppRex AI. Harmonised disease-gene panels from structured clinical and genetic queries. For humans and for machines.
RNNs, causal Transformers, GPT-2-style decoders, self-attention, forward and backward passes, manual gradients, and optimisation in base R.
Hugging Face Transformers, PyTorch, DistilBERT, fine-tuning, task-specific classification, training evaluation, and model validation.
Retrieval, task planning, evidence review, structured extraction, code generation, output comparison, failure inspection, and iterative refinement. Experienced with OpenAI ChatGPT, Google Gemini, and Anthropic Claude.
Retrieval-augmented workflows, embeddings, literature search, mechanistic text summarisation, document chunking, metadata mapping, and source-linked outputs.
Variant interpretation support, report drafting, ICD-10 conversion, candidate diagnosis generation, Molecular Board review, and clinical information structuring.
Open-weight models, Llama, DeepSeek R1, cloud infrastructure, secure high-performance computing, controlled access, and review-ready outputs.
Claim extraction, source tracking, structured reports, manual sampling, validation, audit trails, reproducible code, and reviewable uncertainty.
Customer-facing applications, credentialed users, OpenAI APIs, interface design, workflow design, documentation, stakeholder feedback, and operational rollout.
Rebuilt a recurrent neural network, a causal Transformer, and a GPT-2-style decoder in base R. Implemented attention, forward and backward passes, loss functions, and optimisation directly, without pretrained weights, deep-learning frameworks, or automatic differentiation.
Three reproducible CPU-trained implementations with source code and validation; the GPT-2-style model has 369,504 trainable parameters. RNN · Transformer · GPT-2 · Source code
Fine-tuned DistilBERT using Hugging Face Transformers and PyTorch for four-class news classification. Implemented task-specific training, validation, held-out evaluation, and probabilistic inference.
Held-out accuracy of 88.8% and macro F1 of 0.886 after one CPU training epoch using 1,000 labelled training articles. Methods and results
Built an AI-assisted information layer over structured disease-gene panel evidence, combining UniProtKB mechanistic text with panel metadata, inheritance, disease group, and reported phenotypes.
6.6 million features compressed 5x into disease-aware summaries, integrated into a user-facing application for interpretation and review. Quant Group et al., 2026.
Built a secure AI-assisted clinical genetics system on cloud high-performance computing infrastructure using open-weight models including Llama and DeepSeek R1.
High-throughput variant interpretation and report generation for human review by the Molecular Board, including geneticists, doctors, and researchers.
Developed a credentialed platform using OpenAI APIs and agent workflows to collate clinical information and convert ICD-10 codes into candidate diagnostic hypotheses.
Customer-facing interface, controlled access, structured clinical inputs, candidate diagnosis generation, and reviewable outputs.
Built AI-assisted workflows for systematic review, structured extraction, algorithm development, freedom-to-operate, patent review, and trademark analysis.
More than 5,000 full-text research articles processed into structured datasets; evidence preparation for Switzerland Omics trademark registration with the Swiss Federal Institute of Intellectual Property.
PanelAppRex aggregates disease gene panels and facilitates sophisticated search. Quant Group, Simon Boutry, Ali Saadat, Sinisa Savic, Luregn J. Schlapbach, Jacques Fellay, Dylan Lawless. Bioinformatics Advances (2026). DOI: 10.1093/bioadv/vbag115.
Repository | medRxiv | PDF | Video | Application | About | Dataset
Reply to Dages et al: You AIn’t using it right: artificial intelligence progress in allergy. Dylan Lawless. Journal of Allergy and Clinical Immunology 153(1), 355 to 356, 2024. DOI: 10.1016/j.jaci.2023.09.023.
Building a GPT-2-style language model from scratch in R. Dylan Lawless. Switzerland Omics (2026).
