Start a conversation

Applied AI & LLM systems

Retrieval, agents and evaluation harnesses — built so you can measure quality instead of arguing about it.

  • OpenAI
  • Anthropic
  • LangGraph
  • pgvector
  • Qdrant
  • Ragas
  • FastAPI
  • TypeScript

An LLM demo takes an afternoon; an LLM feature that holds up under real users takes evaluation, retrieval that returns the right chunk, guardrails, and a cost model you can defend to finance.

We build the boring parts that decide whether the interesting part ships: golden sets, offline evals wired into CI so a prompt change is a measurable improvement or it is not, tracing on every call, caching and routing, and a fallback for the day the provider has an incident.

If the feature retrieves, we evaluate retrieval separately before evaluating generation. A model hallucinating and a model never receiving the right chunk look identical in the output and have nothing in common in the fix.

What you get

  • Retrieval pipelines with chunking, hybrid search and re-ranking
  • Evaluation harnesses, golden datasets and regression gates in CI
  • Tracing, prompt versioning and per-feature cost and latency budgets
  • Guardrails, fallbacks and graceful degradation across providers

Common questions

Our prompt keeps changing and we cannot tell if it is getting better. Where do we start?
An evaluation harness, before another prompt edit. Thirty to a hundred real inputs with acceptable outputs, a scorer, and one number in the pull request. It is roughly a day of work and it ends the argument permanently.
Which model provider should we use?
Whichever your evals say, and the architecture should let you switch. We build behind an abstraction thin enough to swap providers without a rewrite, because pricing and capability both move faster than a roadmap.
Can you help with cost and latency?
Yes. Caching, routing cheap requests to smaller models, trimming context, and batching where the interaction allows it. We measure per-feature rather than per-month, because a monthly total tells you nothing about what to change.

Next step

Tell us what has to ship, and by when.

One call, no deck. If we are not the right team for it we will say so, and usually point you at who is.

Start a conversation branislav@thrivee.io

Typical reply within one business day · CET / CEST (UTC+1 / UTC+2)