portfolio
Case studies
AI Fashion App
Computer VisionGenAI3DVirtual try-on for digital and real clothes, built at a startup: pose estimation, garment warping, inpainting with GANs and Stable Diffusion, 3D bodies, cloth simulation and the system around them
Stable Diffusion · Human Pose Estimation · 3D Shape Reconstruction
3D Floorplan Reconstruction
Computer Vision3DTurning an architectural floor plan, scanned, photographed or exported, into a furnished 3D apartment you can walk through in a game engine or a VR headset
Mask R-CNN · U-Net · OpenCV · Unreal Engine 4
Cashierless Store
Computer VisionA store without a checkout: ceiling cameras that follow every shopper on the floor plan, read where they stop and what they take, and send a receipt when they walk out
Multi-Camera Tracking · Object Detection · Camera Calibration · OpenCV
Medical Assistant
NLPBiomedNeuroSymbolic AIAn assistant that reads the medical literature for a doctor: entities found in articles, mapped to medical vocabularies, stored as a graph and ranked by the strength of the evidence behind them
NER · Medical Ontologies · Graph Databases · Language Models
Experiments
Protein Structure Classification
Predicting the CATH architecture of a protein domain, one of ten shapes such as a barrel or a sandwich, from Protein Structure Transformer embeddings: 61% accuracy on superfamilies the model has not seen, where chance is 10%. The structure adds about two points over sequence-only ESM-2.
Protein Language Models · Graph Neural Networks
Floor Plan Detection
Walls and rooms in floor plan images: three MMDetection models fine-tuned on CubiCasa5k, the best a Cascade R-CNN at 0.47 box mAP on the validation split, served by a FastAPI app with a Docker image.
Detection · MMDetection · FastAPI
Geology Forecast Challenge
Forecasting the depth of a geological layer 300 ft ahead of the drill bit from the last 300 ft drilled. A CNN + BiLSTM placed 5th of 197 teams on the private leaderboard of the Kaggle challenge.
LSTM · Time Series Forecasting · Hyperparameter Tuning
Computer Vision Tasks
Four puzzles on rendered images: a shortest route across a terrain map, curved surfaces found with SAM, hidden letters recovered from highlights with DBSCAN, and twelve surfaces grouped into three shapes with k-means.
Path Finding · Segmentation · Clustering
Pandemic Vision
The module allows you to collect statistics and build a dashboard displaying screenshots of violations, as well as elementary graphs for the last 24 hours
Detection · Tracking · Dashboard
Red Russian Gemma
A LoRA adapter for Gemma 1.1 2B-it trained on six books of Stalin's writings in Russian. On a held-out seventh book the loss falls from 4.24 to 2.60 nats per token after 20 minutes on a free T4, and the generations move from assistant-style answers to the register of the corpus.
LLM · LoRA · QLoRA