welcome to my corner of the internet

Liran Attar

Final-year CS student at BGU — starting MSc in CS next year (with thesis).
I like to take complex math, put it into code,
and experiment to see how I can benefit from it.

C++ Deep Learning Machine Learning Quant Finance AI Agents Concurrent Programming
See my work Say hello

Featured Projects

Live System

Garmin Health Intelligence

I wanted to actually use my own Garmin data — not just look at summaries. So I built a pipeline that computes physiological metrics from scratch: Garth data ingestion → PostgreSQL → FastAPI → GitHub Pages dashboard → multi-agent interpretation layer.

The agents are set up so that all the math — HRV baselines, sleep scoring, training load — is computed and validated before any LLM sees it. The agents reason over verified numbers, not raw data. That's the whole idea: keep the math honest.
Python CrewAI PostgreSQL FastAPI GitHub Pages
Case Study & Demo GitHub Repo
Active Project

Quantitative Trading System

A trading system built end to end — the idea, the architecture, the data model, all of it. The Python layer handles IB execution via ib_insync, PostgreSQL persistence, and an async strategy orchestration cycle with per-strategy position isolation and conflict resolution.

The computation-heavy part — scanning for cointegrated pairs using OLS regression and Augmented Dickey-Fuller tests — is a C++ engine compiled and called from Python via pybind11. Strategies: mean reversion and statistical arbitrage on cointegrated pairs. A dry-run mode validates signals without placing orders.
C++ Python pybind11 IB Gateway PostgreSQL asyncio
GitHub Repo
Course Project · Deep Learning

Visual Geo-Localization

Built as the final project for the Deep Learning course at BGU. Given a photo taken somewhere on campus, predict the GPS coordinates — no signal, just the image. Trained an ensemble on a custom 3,646-image dataset with 360° rotational coverage per location.

Architecture: Ensemble of EfficientNet-B0, ResNet18, and ConvNeXt with transfer learning, treating localization as coordinate regression rather than retrieval.

Result: 5.24m mean coordinate error in a GPS-denied environment.
PyTorch EfficientNet ResNet ConvNeXt Transfer Learning
Slides Demo Paper GitHub Repo