AI SYSTEMS ENGINEERINGASTANA · KAZAKHSTAN
I build AI agents and document search. I connect them to your systems and run them on your servers.
From first integration to production operation.
Data, access and control stay inside your perimeter.
Look into the delayed order.
The warehouse moved the shipment to 21 September. Customer: the new date. Logistics: check the dispatch.
Fictional documents for illustration. The order says: “Delivery by 12 October”. The supplier’s offer says: “Delivery by 19 October”. The agent flags the discrepancy; the employee decides on the new terms.
An illustration, not a finished product. Access and approvals are configured to your process.
FOUNDER IN PERSONAleksandr Tikhonov I design, build and support.
EMPLOYMENT EXPERIENCE · LAN E-LIBRARYLanGPT, AI-Snippets AI services for an electronic library.
IN-HOUSE DEVELOPMENTProtocore Agent platform with an open core · beta.
AI services I work on at the Lan e-library, and my own Protocore platform. For each project: the task, my role and the result.
The e-library’s materials become context for the model. The reader gets an answer with references.
The API checks permissions and limits. The request enters a priority queue.
JWT · RabbitMQThe RAG service finds materials and assembles context from source fragments.
The load balancer routes the request to an LLM backend based on its load.
LLM routing · fallbackThe answer reaches the interface as it is generated; references lead back to the sources.
SSE · source markersSimplified backend diagram. Not a screenshot of the LanGPT interface.
Founder’s employment experience, 2024 — present. The figures show the product’s scale, not personal contribution; the role can be confirmed by the employer on request.
More about the systemPer an internal report on a test of one server configuration. Reliability and quality were measured separately; the median does not mean the latency target was met for every request. The interactive example shows the behaviour when generation is off.
More about the systemCore tests and checks are visible in the repository; the platform layer is closed and deployed at the client’s site.
More about the systemMore: Search 2.0, Books Keywords Generator, Daedalus, swiftclf-tuna — all systems and experience.
I help build AI into an existing product or take a prototype to production. Here are problems we can start with.
I bring in the components the task needs. We fix the scope, infrastructure and acceptance criteria before development.
An agent with limited tools, permissions and approval points, working inside the client’s systems.
02Search over the client’s documents with measured quality: hybrid search and reranking select the fragments that answer the question.
03Open models on the client’s hardware via vLLM and Kubernetes; where needed, a hybrid with external APIs under agreed rules.
04Measurement that shows whether things got better after a change of model, prompt or index.
You can order an implementation, bring me in to develop an existing system, or discuss Protocore for your team.
From task analysis to handover into operation: architecture, implementation, deployment inside your perimeter, acceptance on an agreed set.
For a system that already runs: audit, change plan, regression runs, model and index updates, incident analysis.
My own engineering foundation: an open-core agent runtime under MPL-2.0 and a closed platform layer deployed at your site under licence.
Four stages. Each has a written deliverable and an acceptance criterion; the work continues after the model’s first answer.
Data, perimeter constraints, access requirements and quality criteria. Without them there is no way to tell whether things improved.
output: task description, architecture, estimate; written acceptanceRetrieval and prompt variants are compared on a regression set, not on gut feeling.
output: prototype and a metrics table on an agreed setIntegrations and telemetry per an agreed list; monitoring and an operations document.
output: the system in your cluster; scenario-based acceptanceRegression checks on model or index changes, error and latency control, quality reports. The client’s team continues the work independently.
output: reports and an improvement plan; support scope per the contractDesign, implementation and handover into operation: the founder in person. No intermediaries between the problem statement and the code.
Since 2024 I have been responsible for the backend and infrastructure of the Lan e-library’s AI services — LanGPT, AI-Snippets, Search 2.0, Books Keywords Generator: architecture, development, integration, Kubernetes deployment, monitoring and production support. In parallel I develop Protocore, a self-hosted AI agent platform with an open core.
My focus is engineering production systems around LLMs, not training models: services, RAG and backend architecture, integrations, resilience, observability and operation under latency and cost constraints.
IN-HOUSE DEVELOPMENT / ANCHOR INFERENCE
Two products, different jobs. A platform for how a company works and a personal agent that works as a team.
Agents do the work.
You stay in control.
Interfaces, tools and task execution around an open core. For integrating AI agents into a team’s processes.
Open core MPL-2.0
The platform is a proprietary layer, beta.
Works as a team.
Asks all at once.
A personal AI agent with its own window on macOS, Windows and Linux. It hands work to a team of agents, checks every command against a policy and comes to you only with what a person has to decide.
▶ The film “The Task Board” 2:57
macOS · Windows · Linux
Open project MIT · your model and your key
All our capacity is on a current project right now, so we are temporarily not taking new orders. When we reopen, it will show here.
temporarily · not taking new ordersDirectly: hello@anchorinference.com or Telegram @notsoulmate
We are temporarily not taking new orders: all our capacity is on a current project. The request form is switched off until we reopen.
Existing arrangements, contracts and company details: hello@anchorinference.com · Telegram @notsoulmate