In progress · 2026
Genomic Variant Predictor
A small GPT trained on the human genome: does it recognise disease-causing variants?
- AI / LLM
A research project: a language model trained on the GRCh38 reference sequence, whose hidden layers we probe linearly on variants from ClinVar.
Problem
Does a model that never saw clinical labels learn representations that separate benign from pathogenic variants — and does it predict how variants of uncertain significance will later be reclassified?
Solution
Hypotheses written down before results, a frozen dataset described by SHA-256 checksums, sealed test labels with a test guarding against leakage, and a comparison of ClinVar classifications from 2023 and 2026.
- model parameters
- 4.87M
- variants in the main set
- 36,893
- automated tests
- 234
The project is in progress: the data pipeline is complete and frozen, and full training is underway. Results will appear here once the experiment finishes.
Related projects
2026 · Prototype
Anonimizator
Masking personal data before text goes to an AI chat — entirely in the browser.
- AI / LLM
- Security
- Frontend
- Research
2026 · MVP
Olimpio
Competitions and olympiads gathered from the web and email, classified by AI and matched to each student.
- AI / LLM
- Backend
- Frontend
- Data & analytics
2026 · Prototype
Prediction Market Analytics
Prediction-market analytics: wallet scoring, link clusters and a copy-trading prototype.
- Data & analytics
- Backend
- Research