About

I'm a backend engineer building AI systems. Eight years shipping production infrastructure: Spring services and APIs, Kafka and Couchbase for data and messaging, Redis for caching, and breaking a monolith into scalable microservices along the way. More recently I've led a reverse-proxy platform (4 engineers, ~10M requests/day through the fleet, a 75% cut in per-node cost after rebuilding around 2 vCPU / 2 GB nodes and pushing hard on Linux kernel and NGINX tuning). My current focus has shifted to the layer above it, a self-learning debugger that's starting to take routine incident work off the team.

The pivot toward AI / ML happened in the honest direction: I kept building RAG and inference systems at work and noticed that the people shipping them well were backend engineers who had learned enough ML to be dangerous, not ML researchers who had picked up systems. The shortage in this industry isn't people who can fine-tune a model; it's people who can put one in production and keep it there. That felt like a place I could be useful.

Right now I'm going deeper on three fronts:

• Retrieval and knowledge staleness — building aegis-rag as scaffolding to actually benchmark retrieval strategies against each other, and pushing the same ideas into the day-job debugger where fix metadata (commit hash, file references, hit/fail counts) drives an event-driven trusted/stale/quarantined lifecycle for cached knowledge.
• Local inference — running quantized open-weight models on consumer hardware and writing down what works. Eventually this becomes the generation backend for the RAG work. The debugger at work already runs on a Mac Studio with local models; no API bills.
• Computer vision as production QA — training YOLO on custom datasets, both as side work and as one half of the visual-regression QA gate that verifies our deployments before traffic switches. It's the part of computer vision that maps cleanest onto backend intuitions: throughput, latency, failure modes under load.

I don't call myself an AI researcher or an ML scientist. I'm a backend engineer who builds AI systems and is honest about where I am with the ML side. Both halves are load-bearing — backend experience is a real asset for this kind of work, and I'd rather lead with both than hide one.

What I'm shipping

Outside the day job I run Argus Intelligence — a small AI-tooling brand for the EU compliance space. The first product is the EU AI Footprint Scanner, an AST-based static analyser that flags AI/ML library use across a Python codebase and classifies findings into simplified EU AI Act risk tiers. It ships in two shapes: a CLI, and a GitHub App that runs the same analysis on every pull request and posts findings as a Check Run plus a summary comment. Aimed at EU SMEs meeting the August 2026 GPAI obligations without enterprise-grade GRC tooling.

What made the App a real product wasn't the scanner — it was the infrastructure around it. Cloudflare Worker for the webhook receiver, Cloudflare Container running the Python scan, KV-based entitlement gate with a 7-day trial and three paid tiers, Lemon Squeezy webhook for billing, argusintelligence.ee for the marketing and legal surface. Shipping that end-to-end — engine, billing, hosting, legal, GDPR-clean analytics — is the shape of work I want more of.

The Argus name comes from the hundred-eyed Greek guardian — same vigilance metaphor as a continuous PR scanner.

Before this

Before the proxy platform: full-lifecycle client platforms — online auctions, LIMS, HR, POS, RFQ, invoicing — where I first started leading code reviews and raising engineering standards on a small team. And before that, a warehouse management system that hit real e-commerce scale: bin allocation, RFID tracking, Levenshtein matching for replenishment. Both jobs taught me the same lesson: the boring parts break under load faster than the interesting parts.

What's next

Get the day-job debugger to the point where its cache hit rate is compounding — the interesting engineering there is knowledge staleness, not the LLM itself. The first writeup is published; the next one will be numbers on cache hit rate and the failure classes that resist auto-learning.

Grow Argus Intelligence past the first ten paying customers, and turn the PRD for the Argus Compliance Copilot (RAG over the EU AI Act) into a shipped Pro tier. Keep the /now page honest about which of these is actually moving.

Contact

Email: hi@zhen.ee
LinkedIn: linkedin.com/in/zhenee
Résumé: zhen.ee/resume(updated August 2026)