Software thatships.

One platform, three apps, for a distributed interviewer workforce.

An operating system for a call-center network: a Next.js web platform for operations, a Rust/Tauri desktop agent for workers, and a Go media backend that streams live screen and audio for supervision.

• Tauri desktop agent captures screen (H.264) and audio (Opus) and streams it over WebRTC
• Go + Pion backend ingests live streams and records them to HLS
• CATI interview runner, scheduling, chat, documents and performance billing
• RAG assistant over company documents with embeddings and streaming answers

Laundromat automation, from the kiosk to the cloud.

An end-to-end platform for laundromat chains: touchscreen kiosks that top up RFID cards with cash or card payments, an edge backend that drives the machines, and a multi-tenant cloud SaaS where owners run stores, machines, cash collection and analytics.

• Flutter kiosks with RFID readers, bill validators, card dispensers and receipt printers
• Edge backend controls machines over ESP32, with Postgres replication between kiosks for failover
• Kiosks keep working offline and sync transaction batches and telemetry to the cloud
• Multi-tenant SaaS with analytics, cash collection, firmware management and an AI assistant

Storefront, back-office and a Rust realtime server.

Commerce split into three independently deployable apps: a Stripe-powered storefront, a clean-architecture admin dashboard, and a read-only Rust WebSocket server that fans out support chat straight from Postgres.

• Postgres triggers → LISTEN/NOTIFY → Axum WebSockets, zero polling
• Domain / application / infrastructure layers enforced with ESLint
• Fixed-role RBAC with a Redis access cache and 150 ms DB fallback
• Short-lived JWTs authenticated in the first frame, never in URLs

59 million positions per second, written from scratch.

A chess engine and desktop app built to learn Rust properly: legal move generation verified by perft, alpha-beta search with a transposition table, and a Tauri + Preact interface to play against it.

• Iterative-deepening negamax with alpha-beta and quiescence search
• Zobrist hashing with an aging, depth-aware transposition table
• Pins, checks and evasion masks computed once per node with bitmasks
• Packed 16-bit moves and allocation-free attack detection

Ranked 1st at ISIMA — an AI agent that staffs projects from CVs.

A multi-tenant business platform with a voice-enabled, tool-calling AI agent. Managers describe what they need in plain language; the agent creates projects, generates tasks, parses every CV into skills and assigns the best-matched people — with RBAC, tenant isolation and audit logs underneath.

• Ranked 1st and received the Best PFE Project Award at ISIMA
• Tool-calling agent for projects, tasks, users, groups and email — with voice chat
• PDF CV parsing, skill extraction and employee–task matching with confidence scores
• RBAC, tenant isolation, audit logs and Redis caching

Measuring metal hardness from a microscope image.

A U-Net model that finds the corners of Vickers indentations in microscope images to automate hardness measurement — with a custom labelling tool and a FastAPI inference endpoint.

• U-Net trained on a hand-labelled dataset
• Labelling and visual verification tools
• FastAPI endpoint for integration in lab software