Projects
2026
-
01
Uzbek text-to-speech for Openbank
Built the corpus, trained the model and shipped the Uzbek text-to-speech system now serving Openbank's mobile assistant on self-hosted infrastructure. A full fine-tune of OmniVoice on 220.75 hours and 148,529 clips, with leakage-free splits and an Uzbek text-normalisation front end.
-
02
ESM2 to ESM3 representation distillation
An adapter that maps frozen ESM-2 residue representations into ESM3's space, with ESM3 run once offline to cache teacher embeddings for 5,000 UniProt proteins and 1,534,505 residues. On a held-out 500-protein split it reaches R² 0.807 and CKA 0.767 against a raw ESM-2 baseline of 0.691. The success gate was fixed in advance at CKA ≥ 0.80, and it is not met: the adapter predicts ESM3's coordinates well while only partly reproducing its geometry. No distilled checkpoint has been published, because calling this an ESM3 replacement would overstate it.
-
03
BGC-SetNet
A Set Transformer over frozen ESM-2 embeddings for retrieving related biosynthetic gene clusters, evaluated under group-disjoint splits fixed before training. Plain Pfam-domain Jaccard beat it, 0.879 against 0.847 Recall@50, and that is the result I published.
paper (external link)code (external link)model (external link)write-up (external link)
-
04
ESM-2 for protein subcellular localization
Four strategies compared under one harness on DeepLoc 2.0, 24,667 proteins across 10 classes: a linear probe on the 8M model (69.6% accuracy, MCC 0.614), full fine-tunes of 35M (74.3%) and 150M (76.6%, MCC 0.706), and rank-16 LoRA on 650M (76.5%, MCC 0.704) training 2.4% of the parameters. Four checkpoints and an interactive demo are public.
code (external link)models (external link)demo (external link)write-up (external link)
-
05
Uzbek speech recognition
Whisper-medium fine-tuned on the 52-hour FeruzaSpeech corpus to 9.64% word error rate and 1.60% character error rate, with a reproducible pipeline for normalisation, Cyrillic-to-Latin handling and augmentation.