I'm an AI-native engineer: I treat agentic AI (Claude Code, ChatGPT) as my engineering team — I set the architecture, the guardrails, and the review bar; the agents multiply the throughput. It's how one final-year student ships like a small team, without shipping slop:
- 🏭 5+ production full-stack systems for enterprise & government clients in ~12 months
- 🚀 A live SaaS — skipedit.app — solo-built and solo-operated: GPU workers, payments verified with real money in both directions (purchase and refund clawback)
- 🕵️ Adversarial multi-agent bug-hunts on my own releases — one 18-agent hunt caught 18 real bugs (2 critical) before launch; another (24 agents) caught 7 more
- 👁️ An AOI machine-vision system designed, deployed, and validated on a live automotive production line
flowchart LR
A["🧠 Me<br/>architecture · specs · review bar"] --> B["🤖 Agent fleet<br/>build · tests · docs"]
B --> C["⚔️ Adversarial agents<br/>try to break the release"]
C -->|bugs found| B
C -->|clean| D["✅ Me again<br/>final review · ship"]
D --> E["🏭 Production<br/>real users · real lines · real money"]
The loop above is not aspirational — it's the documented workflow behind every project on this page.
| Project | What it is | Stack |
|---|---|---|
| ai-fastener-inspection | Edge machine vision for automotive fastener inspection — 0.995 mAP@0.5, 59 FPS at 15 W on Jetson, validated on a live line. Full ablation + leakage-corrected results | Python, YOLO, TensorRT, Jetson |
| SkipEdit (live) | A SaaS platform that turns lecture recordings + slides into synced presenter-and-slides videos. Solo-built & operated: Next.js, Fastify gateway, Supabase, R2, RunPod serverless GPU workers, Paddle payments | TypeScript, Python, Rust |
| bitnet-rwkv-lm | A 0.15B-param RWKV-7 LLM with BitNet 1.58-bit ternary weights, trained from scratch on an 8 GB GPU — QAT, 2-bit packing, RFT, local chat app | PyTorch, CUDA |
| spiking-ff-jepa | A backprop-free learning study: spiking (LIF) neurons + Forward-Forward + JEPA + int4 QAT — with a full ablation and an honest negative result | PyTorch, Norse |
| kalman-lstm-ppo-trader | An uncertainty-aware RL trader: custom Kalman-LSTM cell feeding RecurrentPPO, with Neural-ODE & FNO ablations | PyTorch, SB3 |
| circuit-idle (play it) | F1-themed idle game — pure, fully-tested TS simulation engine + Three.js dashboard, CI/CD to Pages | TypeScript, Svelte, Three.js |
Fastener-YOLO: accuracy vs. on-device throughput on Jetson Orin Nano — from my funded industrial final-year project.
Work I can't open-source but can talk about all day:
- Machine vision in production — AOI inspection systems developed and validated at customer manufacturing sites in high-value automotive production.
- A five-module enterprise platform for a state agency — portal SSO, per-module RBAC, finance/property/planning modules, and a RAG AI assistant that answers only from the modules each user may open (pgvector semantic search + reranking).
- Industrial IoT — PLC integration over RS485/Modbus, MQTT telemetry into ThingsBoard, Advantech wireless I/O, MIMOS WISP mesh sensors; reverse-engineered a proprietary serial protocol by traffic sniffing; a legacy Firebird accounting DB bridged to a modern REST sync.
AI-assisted development · agentic AI workflows (Claude Code, ChatGPT) · multi-agent adversarial testing · prompt & tool-use orchestration · RAG pipelines ML/DL · PyTorch · quantization-aware training (int4, 1.58-bit) · computer vision · reinforcement learning · TensorRT edge deployment Systems & web · Next.js/React · Node/Fastify · PostgreSQL/pgvector · serverless GPU (RunPod) · CI/CD · Linux VPS ops Hardware · PLCs · RS485/Modbus · MQTT · sensors & real-time control · circuit analysis · Jetson
Most of my day-to-day work lives in private client repos — the public graph is the tip of the iceberg.