I design and ship software that thinks for itself.

Logistics intelligence cockpit

An AI-powered platform that folds real-time fleet tracking, warehouse ops, freight-market intelligence and disruption detection into a single situational-awareness cockpit. A modular Django core streams live data through Redis pub/sub and SSE into a React frontend, with a Gemini-backed AI gateway synthesising narratives and alerts under strict budget controls.

• → Module-agnostic real-time pipeline over Redis pub/sub + Django Channels SSE
• → Pluggable app modules wired through a central bus — no cross-module imports
• → Cost-aware AI gateway with structured Gemini output and budget gating
• → Multi-tenant isolation and HMAC-signed webhooks from day one

On-device form coach, gamified

A fully offline Android app that reads your exercise form in real time using on-device ML Kit pose detection — no internet permission, no cloud uploads. Per-exercise state machines count reps and flag form across squats, push-ups, lunges, planks and jumping jacks, all wrapped in a Solo-Leveling-style "System" of XP ranks, daily quests and achievements.

• → On-device pose detection with zero network permission
• → Debounced form rules + rep-counting state machines per exercise
• → XP ranks (E→S), quests, achievements and a muscle-balance radar
• → Schema-versioned JSON/CSV backups with no storage permission

Reels into a searchable second brain

A FastAPI service that captures Instagram reels shared to a burner account and turns them into a searchable Obsidian vault. It fetches metadata, auto-tags with an LLM, optionally transcribes audio, and embeds everything into a local vector store — then answers natural-language questions across your entire reel library. Every stage is provider-swappable and degrades gracefully when a model is offline.

• → Hands-off reel ingestion via private-API polling
• → Composable RAG: transcription, vision, embedding all provider-swappable
• → LLM auto-tagging into Obsidian markdown with YAML frontmatter
• → Graceful degradation and automatic retry for failed fetches

⚑ Uses a burner account; Instagram ToS risk is acknowledged in the project.

Raw resources to finished decks

A desktop app that ingests PDFs, DOCX, slides, images, YouTube transcripts and web pages, then generates polished presentation decks through Google NotebookLM. A twelve-question interview refines audience and visual style; a seven-section prompt engine drives generation; and a feedback loop banks high-rated prompt fragments so future decks get sharper over time.

• → Multi-format ingestion with automatic research gap-filling
• → Interactive 12-question interview with learned defaults
• → Seven-section prompt engine with visual directives and quality rules
• → Reinforcement loop seeds a success-pattern DB from rated decks

Your phone as a camera FTP dock

An Android app that turns your phone into an FTP server for a Canon EOS R6 (or any FTP-capable camera). Shoot over Wi-Fi, USB-C or a hotspot and watch frames land live in a gallery grid. Swipe to star, reject or skip, then auto-upload keepers to Google Drive, Dropbox or Lightroom — all driven by an embedded Apache MINA server and a WorkManager upload queue.

• → Embedded MINA FTP server in a session-aware foreground service
• → Live gallery grid with swipe review (star / reject / skip)
• → Multi-cloud auto-upload: Drive, Dropbox, Lightroom
• → Wi-Fi, LocalOnlyHotspot and USB-RNDIS network detection

Benchmarking when models should say "I don't know"

An open, frozen-design benchmark for LLM memory systems that measures calibration, not just accuracy. Where other benchmarks ask "did it retrieve the right fact?", GlassBench asks "when it is wrong, does it admit it?" — introducing the Confidently-Wrong Rate and a composite Glass Score across answerable, stale, contradiction and false-premise splits drawn from real multi-session conversations.

• → Frozen pre-registration committed before any system is scored
• → Confidently-Wrong Rate isolates the failures that hurt in production
• → Four task splits across 96 real multi-session conversations
• → Deterministic open scorer that defeats degenerate strategies