Aleksandre Khvadagadze

A multi-threaded Windows service that tracks your hands through MediaPipe, classifies gestures with a custom scikit-learn ensemble, and renders GPU accelerated effects onto a virtual camera device. Any video app reads it as an ordinary webcam feed.

The part I'm proudest of isn't the visuals. The service idles at roughly zero CPU and GPU with the camera off, and wakes only when something actually opens the virtual device. Neither UnityCapture nor OBS exposes an event for that, so I reverse engineered a timing based IPC side channel out of the compiled binary to detect consumers.

Gestures run on an 83 dimensional pose vector built from wrist origin normalization, joint bend cosines, and inter fingertip distances, trained on a 10x augmented dataset. Augmentation runs on the raw landmark arrays before feature extraction, so every derived angle stays geometrically consistent with the pose it came from.

An internal system for a German temp staffing firm that supplies warehouse labour to logistics clients. It handles employee master data, scheduling, attendance, performance, document storage, and an admin only finance ledger.

Ingestion was the hard part. The client ran a decade of work through inconsistent Excel workbooks, so I built idempotent, dry runnable importers that parse attendance, personnel, productivity, and finance files into an Alembic managed Postgres schema. Every row carries provenance, so a bad import can be scoped and reversed without touching manual entries.

Every record is GDPR regulated employee data, which drove the architecture. Hosting stays inside the EU, row level security is on with the public Data API revoked, tokens are verified asymmetrically against Supabase JWKS rather than a shared secret, and PII is masked by role at the schema layer. Finance writes require an admin actor ID, so every mutation has an audit principal attached.

A full product for early stage founders. Generate a brand kit, build landing pages, produce ad creative, and track what competitors are doing, all from a guided onboarding flow.

I wrote 333 of the 509 commits. I owned containerization and the Cloud Build pipeline, Firebase auth, the Firestore data layer, the onboarding and trial verification flow, the ad generation pipeline, and a multi phase interactive product tour with a spotlight host.

The backend runs 11 AI generators covering logos, brand kits, colour palettes, taglines, icons, ads, and page building. Alongside them sits Scout, a market intelligence subsystem with its own HTTP worker that tracks competitors, market motion, and social activity.

A Kubernetes style declarative manifest, pointed at your own life. You write constraints for sleep, workouts, and focus blocks, and a fleet of agents watches reality against them, detects drift, and proposes remediation plans you approve or reject.

The agent pipeline runs orchestrator, then monitoring, then drift detection, then remediation. Gemini 2.5 Flash handles routine passes, and 2.5 Pro gets escalated the conflicting cases.

I built most of the frontend and a large share of the backend integrations, including Google OAuth, Calendar sync, Firebase Cloud Messaging push notifications, ElevenLabs voice briefings, the plan apply modal, and the PWA shell.

A discovery first alternative to searching for food. Users swipe through curated recipe and restaurant cards, and the feed adapts to their reactions, cuisine preferences, and allergies.

I designed and owned the Postgres schema and every API endpoint, 14 route handlers spanning auth, swipes, saves, preferences, recommendations, and place lookup. The backend started as a standalone Express and TypeScript service with five migrations and JWT middleware, which I then migrated into Next.js route handlers when the team consolidated.

Calling Spoonacular on every request would have burned the quota in a day, so I built a daily GitHub Actions pipeline that seeds recipes into the database on a round robin cuisine rotation with rate capping. The recommendation algorithm is frequency weighted with wildcard scoring, so the feed keeps some discovery instead of collapsing into one cuisine.