AI systems that stay inside your perimeter

Agent systems, RAG search and self-hosted inference you can deploy, verify and maintain on your side. Data, models and logs never leave.

Discuss a task

AI systems that stay inside your perimeter

Agent systems, RAG search and self-hosted inference you can deploy, verify and maintain on your side. Data, models and logs never leave.

Discuss a task

Sounding is a stage of the query's route and one of our services. Tap to read.

Anchorage is your perimeter: on-prem or your own cloud.

Hazards are what happens without a perimeter. They open too.

In open sea: press “Plot the course”

Tap a sounding 1–4 or a hazard on the chart: its description appears here.

Sounding is a stage of the query's route and one of our services. Tap to read.

Anchorage is your perimeter: on-prem or your own cloud.

Hazards are what happens without a perimeter. They open too.

the chart is interactive: soundings and hazards open on tapthe ship follows scrolling, hovering or the buttonthat is how we work: everything inside your perimeter

Four directions, one boundary of responsibility

What is included, what we need from you and what you get. The boundary sits where your team makes the decision.

sounding 1

Agent systems

Orchestration, tool calling, memory, permissions and confirmations, prompt-injection defense.

What we need and what you get

From you: process descriptions, access to the systems the agent works with, and a person who owns the rules
Result: an agent with described tool contracts, an action log and confirmation points
Boundary: we answer for the agent's behavior in the described scenarios, the business rules are yours

More about agent systems

sounding 3

Self-hosted inference

vLLM, Kubernetes, closed-perimeter operation, model choice, performance and cost.

What we need and what you get

From you: access to a cluster or GPU servers and your security requirements
Result: a running inference service with monitoring and an operations guide
Boundary: we configure and hand over, hardware and network stay your domain

More about inference

sounding 2

RAG and semantic search

Chunking, embeddings, hybrid search, reranking, retrieval quality evaluation.

What we need and what you get

From you: the document corpus and at least a few dozen questions with expected answers
Result: a search loop with measured quality and a regression set
Boundary: quality is measured on an agreed set, the data stays with you

More about search

sounding 4

Evals and observability

LLM-as-judge, regression sets, metrics, production quality monitoring.

What we need and what you get

From you: access to system logs and a domain expert for labeling
Result: a quality dashboard and regression runs on every change
Boundary: we build the measurement, your team decides on releases

More about evals

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.

The system does not end at the model's answer

Four stages. Each has a clear boundary and a written result at the end.

Task breakdownweek 1

We pin down what to measure and how

Data, perimeter constraints and quality criteria. Without them nobody can tell whether it got better.

output: task description and estimate

Prototype with metricsmeasurable

A first version on your data

Retrieval and prompt variants are compared on a regression set, not on gut feeling.

output: prototype and a metrics table

Deployment in your perimeteron-prem or cloud

vLLM and Kubernetes in your infrastructure

Models, embeddings and logs never leave the perimeter.

output: the system in your cluster

Operation and evalsongoing

Production quality under watch

Regressions on model changes, prompt-injection defense, quality reports.

output: reports and an improvement plan

Open code instead of promises

The core we build agents on is public. You can read it, run it and check it before the first call.

# a core for stateful LLM agents, MPL-2.0 $ pip install protocore # reason-act loop, tool contracts, state carried between steps $ git clone https://github.com/anchor-inference/protocore-community # an autonomous agent in a Linux container on top of Protocore $ git clone https://github.com/ascorblack/daedalus no invented clients or numbers: only code and tests

protocore

A protocol-first core: every step is described, every tool call is verifiable. Python, asyncio.

github.com/anchor-inference/protocore-community

daedalus

An agent living in a container: tools, planning, limits on what it may do. Changes to its own code only through a reviewable pull request.

github.com/ascorblack/daedalus

employment

An electronic library system: large text corpora and semantic search over them. Internal architecture and volumes are not published, they belong to the employer.

A contract with a company, not a hire on staff.

The contractor is a Kazakhstan LLP. Contracting it does not require a foreign-labor permit, unlike hiring a foreign national on staff. Code ownership is fixed in the contract.

this is not legal advice, check with your counsel

What happens after you write

one business daywe reply and offer a call slot

30 minutesa short call about the task and the perimeter

in writingtask description and estimate

Directly: ascorblack@gmail.com or Telegram @notsoulmate

Full form and what we need for the first call