01services · ai.architecture

AI architecture. Costed, evaluated, shipped.

Retrieval that refuses to answer without a source. Models routed by cost, not by habit. Evaluation that catches the regression before your customer does. Designed against your data and your budget, not a vendor's slide deck.

02the problem

The demo worked. Production is a different system.

A RAG pipeline that retrieves the wrong paragraph does it with total confidence, and the model downstream will dress that paragraph up as an answer. A single strong model wired to every request works fine until the invoice arrives. And a prompt change that quietly breaks one workflow in ten will not announce itself: without an evaluation suite, your customers are the test harness.

None of this is exotic. It is the ordinary engineering of AI systems, and it is decided at the architecture stage: how documents are chunked and retrieved, when the system must say "I don't know" instead of guessing, which model tier handles which class of request, and whether your data can leave your infrastructure at all.

03what I do

Four disciplines, one design.

Each engagement produces a written design with the tradeoffs stated, and where the shape calls for it, working code.

ai.01

RAG & retrieval design

Chunking, embedding, hybrid search, and the part most pipelines skip: a grounding guard. If no source clears the relevance threshold, the system escalates to a human instead of improvising. Postgres + pgvector where it fits, managed vector stores where it genuinely pays.
ai.02

LLM routing & cost control

Cost and latency budgets modelled before the prompt is written. Cheap models for classification and routing, strong models only where the margin justifies them, and an adapter seam so switching providers is a config change, not a rewrite.
ai.03

Evaluation & hallucination control

Regression suites for prompts and retrieval, so a change is measured before it ships. Where a wrong number is unacceptable, the design computes it deterministically and never lets the model guess.
ai.04

Self-hosted vs managed

GDPR posture, data residency, cost curves at your actual volume, and operational burden, written down as a tradeoff table you can defend to a board or a regulator. Sometimes the answer is managed. The point is deciding, not defaulting.
04proof

Systems that refuse to guess.

Ospite is an AI concierge platform for hospitality venues: FastAPI + Next.js + Postgres/pgvector, with tenant isolation enforced by row-level security in the database rather than by application-code discipline. Its retrieval layer carries a grounding guard: when no source chunk clears the relevance threshold, the request escalates to a human. The architecture routes cheap models to classification and strong models to complex turns, behind one adapter seam.

Verax is a compliance intelligence platform for the iGaming vertical: AML/KYC.

The same discipline runs through the MCP work: auseklis exists because a language model asked for a planet's position will happily invent one. Twelve tools compute the real answer from a real ephemeris, and the model is never allowed to do the math. That rule (compute what must be correct, generate only what may be wrong) is the core of every AI architecture I ship. More on the protocol side at MCP consulting.

If your AI feature works in the demo and worries you in production, that's the conversation. One email, one business day to a reply.

Start a conversation →gmz@gmz.lv