Production AI specialist. Open to freelance.
RAG and AI agents that survive production.
I build production RAG and AI agents, and I cut what they cost to run. Most AI never survives the demo, the hard part is drift, latency, evals, and cost. I shipped a live system that handles all of it. Go click it.
0
AI tools
~0%
cache hit
0+
KB tables
0+
build cycles
What I specialize in
The scarce, high value work. Each lane is proven by a real, shipped feature in Agrotus.
Production RAG over your data
An assistant that answers from your own documents and databases, grounded so it does not invent answers.
RAG over an 80 plus table knowledge base with grounding guards, live in production.
AI agents that run in production
Tool using agents that actually do the work and keep working under real load, where most agent demos die.
A 45 tool agent running live, not a proof of concept.
Cost control, evals, and observability
Cut what your AI costs to run and keep it reliable, with caching, routing, evaluation, and drift checks.
About 95% cache hit, about 4 cents per answer, 50 plus build cycles with adversarial QA.
AI built into your existing software
Models integrated into a real product and shipped. Backend, auth, deployment, third party data.
A full stack platform with satellite and vision integration, deployed solo on a self managed server.
Custom models when off the shelf will not do
A trained model for your domain when a generic LLM cannot do it.
A custom trained vision classifier.
Most AI systems never leave the demo stage. The hard part is everything after: drift, latency, evals, cost, and keeping it reliable. That is the part I have actually shipped.
The proof
Agrotus
Challenge
Build a production AI platform for farming end to end, solo. Grounded agronomy that does not hallucinate, real satellite and vision pipelines, multilingual, secure, and cheap enough to run on one small server.
Solution
A Claude powered assistant with 45 tools over an 80 plus table knowledge base, a custom trained disease model, Sentinel 2 analysis, yield forecasting, and a full operations cockpit. One streaming call per message, dual breakpoint caching, hallucination guards.
Result
Live in production. About 95% cache hit, about 4 cents per answer, EN and LT. Designed and shipped solo in about 3 months, across 50 plus build cycles.
Which part proves which niche
How it is built
Enough specifics to trust the work. Verified against the live codebase.
AI
Manager agent loop with typed tool dispatch over Claude Sonnet. 45 tools, dual breakpoint prompt caching at about 95% hit, one streaming call per message, hallucination guards that keep numeric and regulatory answers grounded in the database.
Cost control
Split conformal routing and token budgets bound model spend to a few cents per message.
Vision and ML
A custom trained DINOv2 nine class crop disease classifier, plus a multi signal photo diagnosis path live in the product.
Geospatial
Sentinel 2 imagery to NDVI to K means zones to ISOXML export.
Data
An 80 plus table knowledge base with retrieval at inference time.
Security
JWT in httpOnly cookies, CSRF tokens, organization scoped queries, a read only public demo enforced by middleware, rate limiting.
Stack
Next.js, React, TypeScript and Tailwind on the front. FastAPI, SQLAlchemy, Pydantic and Python on the back. PostgreSQL, SQLite and pgvector for data and retrieval. Self managed VPS with systemd and nginx.
Let’s build something.
Tell me what you are building. I can integrate an AI assistant on your data, train a model for your domain, or build the whole system. Fixed scope, shipped. Just the engineer who does the work.
Available for freelance. AI integration, ML models, or full builds.
Designed and shipped solo in about 3 months, across 50 plus build cycles.