A boutique practice for AI in sequence biology.
Pragmatic advisory and empirical analysis for biotech, pharma, and research teams putting machine learning to work in genomics, proteins, and drug discovery.
The field is moving quickly, and much of it is oversold. Generalised models that look state-of-the-art on clean academic benchmarks often degrade when they meet domain shifts, messy clinical data, and real hardware constraints.
rewire.it exists to navigate that gap. The work is hands-on, independent, and candid about where machine learning helps and where a simpler approach is the honest answer. The writing on this site is a fair sample of how the practice thinks.
What we help teams with
Strategic feasibility & architectural selection
A reality check before the spend: whether a multi-billion-parameter foundation model is genuinely required for your problem, or whether a smaller, task-specific architecture would perform better. Covers hardware constraints, inference cost, open versus proprietary models, and tokenizer choice for sequence data. Delivered as reviews, workshops, and written briefs.
Pipeline architecture & evaluation design
Rigorous benchmarking protocols and genuinely independent, held-out test sets to measure the clinical or experimental viability of your models — surfacing domain shifts, overfitting, and information bottlenecks before they reach production. Usually project-based.
Specialised fine-tuning & capability strategy
For teams sitting on proprietary data: how to adapt and fine-tune foundation models for specific sequence-biology tasks — variant-effect prediction, design, annotation — without misapplying language-model assumptions to biological sequence data.