Home/Services

Four directions, one boundary of responsibility

Each direction is described the same way: what is included, what we need from you, what you get and where the boundary lies. These are the same four soundings as on the chart on the home page.

sounding 1

Agent systems

An agent that does useful work in your systems and never steps outside the described rules. Every tool call is described by a contract, and a human confirms risky actions.

What is included

  • orchestration and the reason-act loop on the Protocore core
  • tool contracts: what the agent may call and with which permissions
  • memory and state between steps, recovery after a failure
  • human confirmation points for risky actions
  • prompt-injection defense and a log of every action

What we need from you

  • descriptions of the processes the agent should run
  • access to the systems it works with: database, API, documents
  • a person on your side who owns the rules

Result

  • an agent with described tool contracts
  • an action log and confirmation points
  • scenarios the behavior is checked against

Boundary of responsibility

We answer for the agent’s behavior in the described scenarios. Business rules and access rights are set by you.

sounding 2

RAG and semantic search

Search over your documents with quality that is measured, not “seems fine”. Hybrid search and reranking select the fragments that actually answer the question.

What is included

  • chunking and corpus preparation
  • embeddings and a vector index inside the perimeter
  • hybrid search: lexical plus semantic
  • reranking and noise filtering
  • retrieval quality evaluation on a regression set

What we need from you

  • the document corpus in any format
  • at least a few dozen questions with expected answers
  • an expert who can say which answer is correct

Result

  • a search loop with measured quality
  • a regression set you can rerun on every change
  • a guide to updating the corpus

Boundary of responsibility

Quality is measured on an agreed set. Data and the index stay in your perimeter.

sounding 3

Self-hosted inference

Open models on your hardware through vLLM and Kubernetes. No external APIs: models, queries and logs never leave the perimeter.

What is included

  • vLLM in Kubernetes or on dedicated servers
  • choice of open models for the task and the hardware
  • performance: batching, quantization, cost per request
  • monitoring and alerts
  • an operations guide for your team

What we need from you

  • access to a cluster or GPU servers
  • security and network requirements
  • an idea of the load: how many requests and what latency you expect

Result

  • a running inference service with monitoring
  • a guide to operating and updating models
  • performance and cost measurements

Boundary of responsibility

We configure and hand over. Hardware, network and access stay your domain.

sounding 4

Evals and observability

Measurement that shows whether things got better after every change: a new model, prompt or index. Degradation is visible before a user sees it.

What is included

  • regression sets on your data
  • LLM-as-judge checked against humans
  • quality metrics and a dashboard for the team
  • production quality monitoring
  • runs on every change to the system

What we need from you

  • access to system logs
  • a domain expert taking part in labeling
  • criteria for what counts as a good answer

Result

  • a quality dashboard
  • regression runs on every change
  • reports and an improvement plan

Boundary of responsibility

We build the measurement. Your team decides on releases.

Stack: Python, asyncio, FastAPI, PostgreSQL, vLLM, Kubernetes. Only what we actually use. We do not publish prices: every engagement is estimated after the task breakdown. Describe the task and we will offer a call slot.