Projects / Проекты

Project archive

Overview / Обзор

Machine learning, computer vision, GPU computing, and developer tools.

5 projects shown.

Featured project

2025 · Shipped

EV-Pulse

Forecasts the power demand of an EV charging site 24 hours ahead from calendar and weather context, so facility managers can anticipate demand peaks instead of tripping breakers or paying capacity penalties.

ML

Context
Each site is modelled separately from the ACN charging dataset: a recent reference profile (the mean of the last eight training weeks for the same weekday and time slot) plus a LightGBM residual correction driven by calendar and weather context. Served by FastAPI with a Streamlit dashboard, packaged with Docker Compose.
Outcome
On a strictly held-out 60-day window, MAE drops to 4.78 kW at Caltech (29.8% better than the calendar baseline) and 5.35 kW across all sites (14.7%). Where the residual model fails to beat the baseline, the API deliberately keeps the baseline instead.
The Signs home screen: lesson 4 of 12 of the LSF alphabet course, with a start button and a photograph of a hand forming a letter.

02

2026 · Shipped

Signs

A computer vision project for learning the French Sign Language alphabet. It recognizes letters from your webcam in real time, with every frame staying on your machine. Evaluation focuses on generalization to new users.

MediaPipe locates 21 hand landmarks, which are re-centred on the wrist and scaled by palm size so the model sees a shape rather than a position, then a 147 KB ONNX MLP classifies the letter and a 9-frame rolling average stabilises it. All of it runs in the browser tab; the Django REST backend is optional and only stores progress once you sign in.

90.2% top-1 accuracy on an unseen signer, at ~30 fps on a commodity laptop. The finding that mattered: hand-crafted geometric features scored best on the conventional same-signer metric but cost at least 4 points on a new person; optimising that metric would have shipped the wrong model.

React 19 · ONNX Runtime Web · MediaPipe · scikit-learn · Django 5

A deep zoom into a Julia set rendered by FractalFlow: golden spiral filaments studded with miniature copies of the set, against a deep blue field.

03

2026 · Live demo

FractalFlow

An interactive Julia set explorer that zooms to 10⁻²⁸ in the browser, far past the ~10⁻⁵ limit of float32, using perturbation theory with a double-double reference orbit.

One algorithm implemented three times: WGSL (WebGPU), GLSL (WebGL2 fallback) and CUDA C, then cross-validated pixel-for-pixel against an arbitrary-precision Python renderer, which caught a real rebasing bug during development.

Sustains ~0.9 trillion iterations per second in the browser on an RTX 3060 Ti, roughly 40,000× a vectorized NumPy baseline. Every view is shareable as a URL.

TypeScript · WebGPU / WGSL · WebGL2 / GLSL · CUDA C · mpmath

04

2024–2026 · Ongoing

Kaggle Competition Workflows

A continuing body of competition work focused on reproducible pipelines, feature engineering, ensembles, and disciplined evaluation.

A working ground for tabular modeling, fast iteration, and leaderboard-sensitive decisions.

Keeps experimentation tied to measurable leaderboard results.

Python · PyTorch · scikit-learn · Notebook workflows

The Caraml browser IDE running an OCaml program, with the editor on the left, a memory panel showing the environment and heap on the right, and evaluated output in the console below.

05

2026 · Live demo

Caraml

A full OCaml IDE in the browser: Monaco editor, multi-file projects, code execution, and a live visualization of the stack, heap, and environment.

A React + Express app with SQLite that I built to make OCaml approachable in the browser. It uses the system OCaml toolchain (merlin for completion and types, ocamlformat) when installed, and falls back to an in-browser OCaml interpreter otherwise. Integrates with Learn OCaml for exercises and grading.

Students can write, run, share, and get graded on OCaml code with nothing installed on their machine.

TypeScript · React · Express · Monaco Editor · OCaml