Kolosal AI
Open-source platform for running large language models locally, so your data stays private and inference stays fast. Custom training and production-ready inference at scale.
PythonPyTorchFastAPIDockerCUDA
Evint Leovonzko
Cofounder of
Kolosal AI
Making AI genuinely useful, at Kolosal AI and previously at Genta Technology. Researcher at UBC, where I implement and explore how machines decode biology (from circular RNA to tumor suppressor genes) and, occasionally, the stock market.
Recent activity
Things I've built
Open-source tools, platforms, and applications
Visit Platform
PyTorch Inference Framework
Production-ready inference with TensorRT acceleration. Handles batching, model versioning, and GPU memory management. Built under Kolosal AI.
PyTorchTensorRTFastAPIDockerCUDAPythonKolosal AutoML Platform
End-to-end AutoML with a Gradio UI: from hyperparameter tuning with Optuna to experiment tracking with MLflow, all in one open-source platform built under Kolosal AI.
GradioMLflowOptunaFastAPIPythonscikit-learnSOM Plus Clustering
An extended self-organizing map library with smarter initialization options and built-in cluster quality metrics, so unsupervised topology learning is more rigorous and reproducible.
PythonNumPySciPyscikit-learnMatplotlibAPT Fitness Assistant
Real-time pose estimation with MediaPipe for form correction, rep counting, and personalized workout recommendations: your AI gym partner, running locally from your webcam.
PythonMediaPipeOpenCVStreamlitNumPyChain Reaction Simulation
A high-performance nuclear chain reaction simulator written in Rust: GPU-accelerated particle physics, real-time 3D visualization, and a 150+ isotope library drawn from ENDF/B-VIII.0 nuclear data.
RustwgpuwinitVulkan / Metal / DX12ENDF/B-VIII.0Faction: Stock Analysis App
A desktop app for evaluating stocks through technical and fundamental analysis, powered by on-device ML inference and a local Qwen2.5 1.5B LLM. No cloud, no subscriptions, fully private.
Tauri v2RustReactTypeScriptcandle (HuggingFace)Qwen2.5 1.5BMetal / CPURebirth Educational Platform
A community-driven platform connecting learners and researchers through structured courses and peer mentorship, built on the belief that good education should be accessible, not gatekept.
PsychIDN Community
Building mental health awareness across Indonesia through accessible psychology content and peer-to-peer support networks, because understanding your mind shouldn't require a degree or a therapist's fee.
TRIBE v2: Neural Prediction Workstation
An interactive 3D brain visualization tool that predicts cortical activation patterns across six regions in response to text, image, and audio stimuli. It runs LLaMA, CLIP, and Wav2Vec encoders locally in Rust.
RustThree.jsHTML / CSS / JSLLaMACLIPWav2VecPythonorderflow-rs
A high-performance Rust pipeline for limit order book microstructure research. From a single binary, it ingests raw tick data, computes order-flow imbalance features and 87 technical indicators, runs synthetic LOB simulations, and backtests signal-based strategies.
RustParquet / ArrowSpearman ICLOB ReconstructionOFI FeaturesWalk-forward BacktestCriterion Benchmarks
Papers & analyses
First-author research in ML, computational biology, quantitative finance, and computational economics
Fiscal Stabilisers, Minsky Dynamics, and Distributional Outcomes in a Keynesian Agent-Based Model
A daily-frequency, single-sector Keynesian ABM of 10,000 households, 1,000 firms, and 10 commercial banks reproduces four stylised macroeconomic facts and documents persistent zero lower bound binding, a novel result in the K+S ABM literature mirroring post-1998 Japan and post-2013 Euro-area experience.
Individual Optimality and Collective Failure: Survival-Maximising Strategies in a Keynesian Agent-Based Model
Grid Search over Agent Decision Rules Reveals a Structural Coordination Failure Attractor. Using a grid search over key decision parameters for households, firms, and banks, this paper identifies survival-maximising rules empirically, without imposing theoretical equilibrium conditions. When all agent types simultaneously follow their individually-optimal strategies, aggregate outcomes diverge sharply: GDP growth falls 65%, firm bankruptcies rise 17-fold, and mean firm profit turns negative.
SOM-TSK: Topology-Seeded Clustering Framework
SOM-TSK exploits the manifold-mapping properties of a trained Self-Organizing Map to generate topology-guided seed pools for deterministic K-means: it matches or exceeds KMeans++ on every one of 24 benchmark datasets, with zero losses.
GRASP: Graph-Routed Adaptive Spectral Partitioning
GRASP is a parameter-free clustering pipeline that constructs a topological fingerprint of any dataset, estimates the natural cluster count automatically, and dispatches to the most suitable specialist algorithm without requiring labels, validation data, or user-specified hyperparameters.
Dual-Encoder vs Cross-Encoder for Transductive GRN Link Prediction
A rigorous parameter-matched comparison of dual-encoder and cross-encoder architectures for gene regulatory network link prediction: ablation studies, pruning experiments, imbalance robustness, and cold-start evaluation.
Modular vs Monolithic Architectures for GRN Edge Prediction
Gradient Stability Analysis and a Controlled Cross-Architecture Comparison. A controlled comparison of modular two-tower models against monolithic cross-encoders for GRN inference. We diagnose three critical gradient failures in the two-tower design and demonstrate that the cross-encoder outperforms it, especially under class imbalance.
Two-Tower Networks for GRN Inference
A pure-Rust two-tower MLP that learns entity embeddings and cell-type expression profiles to predict transcription factor–gene interactions: 83% ensemble accuracy, CPU-trainable without any deep learning framework.
Deep Learning for Circular RNA Classification
A lightweight ANN pipeline that turns circAtlas k-mer frequencies into reliable disease predictions: fast enough for real-time screening, accurate enough for clinical relevance.
Functional Group Analysis of Drug-Like Chemical Space
Graph-level variational encoding, stratified unsupervised clustering, and formal enrichment testing map how functional group composition varies across 249,455 ZINC15 drug-like molecules, with counterfactual QED analysis decomposing scoring artefacts from genuine chemical signals.
GDP Trajectory Clustering
190 countries, 45 years, four distinct development trajectories. Machine learning clustering reveals why some economies surge while others stagnate, and what policymakers can do about it.
Crypto vs Stock Timing
Do equity markets still lead recessions, or has crypto changed the playbook? Empirical analysis of cross-asset dynamics across five business cycles.
Bitcoin Portfolio Allocation Analysis
Optimal BTC sizing via Risk-Budget Framework: Component Risk Contribution analysis across five portfolio profiles. The answer is always between 0% and 16%.
What Does the Internet Do to the Brain?
Activation Cartography maps 3,008 natural language stimuli across 13 internet content categories against predictions from TRIBE v2 (a 177M-parameter deep neural encoder trained on real fMRI recordings), revealing statistically significant, category-level differences in predicted cortical recruitment.
Cancer Leading Mutation DNA of P-53 Gene
Closed-loop genetic algorithms surface the earliest mutation signatures that destabilise P-53 (the "guardian of the genome") before malignant cascades take hold. Award-winning research that connects evolutionary computation and cancer genomics.
03 / Background
Experience
- 2022 to 2026
University of British Columbia
Undergraduate Researcher
Building deep learning pipelines for circular RNA classification and genomic sequence analysis.
- 2023 to Present
Kolosal AI
Co-Founder
Architecting open-source tools for local LLM deployment, inference optimization, and MLOps automation.
- 2024
Publication
First Author
“Deep Learning Algorithm with Gaussian Blur Data Pre-processing in Circular RNA Classification”
About
I chase problems where data hides something meaningful: a motif buried in RNA sequences, a leading signal in financial time series, or a chance to make ML infrastructure less painful.
Currently an undergraduate researcher at UBC building deep learning systems for genomic data, and co-founding Kolosal AI to make running LLMs locally actually simple. My first-author publication on circular RNA classification is live in URNCST Journal.
When I'm not debugging loss curves, I'm reading about macroeconomics or exploring Vancouver's trails.
More about meWhat I work with
- Python
- PyTorch
- TensorFlow
- FastAPI
- Docker
- CUDA
- Pandas
- OpenCV
- Streamlit
- MLflow