Projects
Research collaborations, end-to-end ML systems, bioinformatics tools and experiments.
2D → 3D Reconstruction — VistaDream + Depth Anything
researchIntegrated Depth Anything v2 into the VistaDream pipeline via a custom adapter. Modified depth normalization and camera logic for stable 3D projection from single 2D painting images. Generated 3D Gaussian Splatting (.ply) point clouds enabling interactive 3D exploration of museum artworks.
ML Digital Painting Restoration — Diffusion + LoRA + FLUX
researchDiffusion-based inpainting pipeline (FLUX.1-dev) for restoration of damaged historical artworks. Trained LoRA adapters for painter-specific style reconstruction. Improved stability via mask refinement and controlled inference tuning.
Task 1 — ChEBI Ontology Prediction
drug designHierarchical multi-label classification of small molecules into 500 ChEBI ontology classes. Used molecular fingerprints (ECFP, MACCS), GNNs and SMILES transformers. Classes form a DAG — child predictions must be consistent with parents. Metric: macro-averaged F1.
Task 2 — Context Retrieval for AI Coding (JetBrains)
LLM / RAGBuilt a precision Context Collection Pipeline for Python codebases in collaboration with JetBrains Research. The pipeline retrieves the most relevant code snippets for a given developer query, used as context for LLM code completion. Evaluated via ChrF score across three models: Mellum, Codestral and Qwen2.5-Coder.
Task 3 — Heat Pump Grid Load Forecasting (Euros Energy)
time-seriesTime-series extrapolation of electrical grid load for a network of heat pumps across Poland. Trained on 7 months of 5-minute telemetry (13 temperature sensors + geolocation metadata) to forecast per-device monthly load for May–October 2025. Extrapolation task: model must generalize to a new year. Metric: MAE.
Task 4 — ECG Paper Digitization
medical CVComputer vision pipeline converting scanned or photographed paper ECG printouts back into 1D digital time-series signals at 500 Hz. Handles real-world degradation: crumpled paper, handwritten annotations, B&W photocopies and poor lighting. Scored across three axes: signal morphology (Pearson correlation), amplitude calibration (SNR) and temporal alignment (cross-correlation).
ADMET Multi-Task Learning
donePredicting full ADMET drug profiles (Absorption, Distribution, Metabolism, Excretion, Toxicity) using a multi-task learning architecture. Shared representations across pharmacokinetic endpoints improve data efficiency and generalization.
ECGApp — Full-Stack ECG Monitoring & ML Classification
flagshipEnd-to-end ECG system spanning hardware, mobile and ML. Arduino acquisition → BLE → iOS app (SwiftUI + CoreBluetooth) → three independent PyTorch models for rhythm classification, beat-type classification and waveform segmentation. Deployed on-device via CoreML.
ECG Arrhythmia Classifier
Deep learning model for arrhythmia detection from raw ECG signals. Preprocessing: resampling, wavelet denoising, normalization.
ECG QRS / P / T Wave Recognition
Waveform segmentation model identifying QRS complexes, P waves and T waves in ECG recordings.
ECG Signal Generator
Synthetic 10-second ECG signal generator for testing and training data augmentation.
DNA Processing Pipeline
Pipeline for DNA preprocessing and variant calling on genomic data in VCF/BAM format using GATK and samtools.
GeneticViewer
Desktop application for visualizing and exploring genetic data files.
FoxyGodot — 2D Platformer Game
2D platformer in Godot (C#) designed as an exercise in classical OOP design patterns. Implements Strategy, State, Factory, Observer / Event Bus, Singleton and Template Method.
FlappyBirdGodot — 2D Flappy Bird Clone
Flappy Bird clone built in Godot with C#. Features scene management, signal-based event system (SignalManager, GameManager), score tracking and animated transitions. Architecture uses partial classes and Godot's node lifecycle (_Ready, _Process, _ExitTree).
LLM Multi-Agent Experiment
Multi-agent experiment using llama-dolphin models exploring agent communication, role-playing and adversarial prompting dynamics.
FlashMind — Flashcard App
Flashcard learning application.
// more on github.com/FBI223