Home/Systems and experience

Systems and experience

Systems taken to production, with precise attribution: what was done as an employee at the Lan e-library and what is in-house development. Figures are given with their source and measurement conditions.

Founder’s experience at the Lan e-library (2024 — present)

Backend and infrastructure for the AI services of an educational platform: architecture, development, integration, deployment, monitoring and production support. Stack: FastAPI, Python 3.12+, Kubernetes on a GPU cluster, Elasticsearch, RabbitMQ, Redis, vLLM, Prometheus.

production · pilot results published

LanGPT — a RAG/LLM assistant for academic work

An assistant integrated into the Lan e-library platform: formulating hypotheses, choosing a methodology, summarization, sections with citations, checking academic content. It answers from verified sources in the e-library catalogue, available to the company under partnership agreements with publishers, without access to the external internet.

my area of responsibility

  • public API: JWT sessions, rate limiting by plan, SSE streaming, task queuing
  • a RAG service on top of Elasticsearch and a pipeline of more than ten academic writing tools with request moderation
  • LLM backend load balancer: dynamic server registry, least-loaded / least-connections with priority, fallback, OpenAI-compatible adapters
  • observability: Prometheus, Grafana, structured logs and analytics in Elasticsearch
  • deployment via Helm to Kubernetes on GPU infrastructure

result

Pilot results published by Lan on 17 April 2026: 20 universities, 6,500 users, more than 37,000 queries. Public source.

The figures show the product’s scale, not personal contribution; the role and area of responsibility can be confirmed by the employer on official request.

Python · FastAPI · PostgreSQL · MySQL · RabbitMQ · Redis · Elasticsearch · vLLM · Prometheus · Grafana · Kubernetes · Helm · GitLab CI

production · load test

AI-Snippets — answers above search results

The service shows a short snippet answer with source markers above the search results. If something fails, the snippet simply doesn’t appear — search and results work as usual.

solution

  • pipeline: two-level cache (exact + semantic) → query classification → expansion → multi-source search over Elasticsearch → context assembly with BM25 + RRF → generation on vLLM → quality gate
  • polling via Redis, distributed semantic cache, circuit breaker, graceful degradation
  • tuning vLLM parameters for performance on a specific server configuration

measurement

13.2 hof continuous load
55 200requests
0infrastructure errors (crashes and 5xx)
87,3 %passed the quality gate; the rest degraded gracefully with no error surfaced to users
4.6 smedian latency against a 5 s target

Per an internal load test report on one GPU server configuration. Infrastructure reliability and answer quality were measured separately; the median latency does not mean the target was met for every request.

Python · FastAPI · RabbitMQ · Redis · Elasticsearch · vLLM · sentence-transformers · Prometheus

Books Keywords Generator — catalogue enrichment

Two-stage LLM generation of book keywords from the table of contents: the first model proposes candidates, the second validates them for language, length, uniqueness and meaningfulness. Dynamic context length control, a concurrency limit to protect the GPU, structured JSON logs in Elasticsearch; generation is skipped for books with a poor-quality table of contents. The results are used in MARC records and search.

Python · MySQL · vLLM · Elasticsearch · Docker

In-house development

Developed outside working hours and without employer resources, under the anchor-inference organization on GitHub. The open Protocore core is the foundation of the agent systems Anchor Inference builds.

in-house development · founder / solo developer · beta

Protocore — a self-hosted AI agent platform

An AI agent orchestration platform for organizations that cannot hand data to external SaaS: banks, the public sector, education, industry. It closes the gap between cloud AI services that require handing over data and frameworks from which the product has to be assembled by hand. Deployed in the client’s infrastructure; supports vLLM, Ollama and OpenAI-compatible providers.

runtime and orchestration

  • an agent loop with limits on tokens, iterations and tool calls
  • multi-agent orchestration: subagent strategies, structured handoff, context compression for long sessions
  • skills and plan tracking; a tool sandbox with capability policies and multi-step approval of dangerous operations
  • extension through typed contracts, hooks and execution strategies; the core is not tied to adapters

enterprise backend and security

  • FastAPI HTTP API with SSE streaming, a scope-aware control plane and an operations panel
  • RBAC with six scoped roles; access plans with quotas, rate limits and violation policies
  • LLM routing with health monitoring, retry, circuit breaker and quotas; a versioned control plane in PostgreSQL with publish / rollback
  • RabbitMQ with priorities and dead-letter, OpenSearch for runs / traces / logs with retention; an autonomous worker with leader election

platform architecture

frontendsprotocore-chat · protocore-dashboard — chat and admin panel
supportingtool sandbox · embeddings · autonomous worker
enterpriseFastAPI · control plane · LLM routing · RBAC · access plans
core runtimeprotocore — contracts, orchestration loop, hooks, safety, observability · open core, MPL-2.0
deployHelm charts · environment profiles · CI/CD — release boundary
dependencies: frontends and supporting → enterprise → core

status

The open core is published under MPL-2.0: installation instructions and the current pre-release are in the repository protocore-community with tests and CI checks. The platform layer is closed and deployed at the client’s site. Stage: beta; deployment scope and licensing terms are discussed separately.

stack

Python 3.12 · FastAPI · Next.js 15 · React 19 · PostgreSQL · Valkey (Redis) · RabbitMQ · OpenSearch · Helm · Kubernetes · Harbor · GitLab CI

Protocore product site

open project · MIT

Daedalus — a personal AI agent

A personal agent with its own window on macOS, Windows and Linux, with or without Docker; on a server it is also reachable from a phone and, if you want, in Telegram. Every project has a team of agents — Daedalus agents or CLI agents such as Claude Code and Codex; questions for the person collect and arrive all at once. Every tool call passes a policy: allow, ask or deny. The daily spending cap is set outside the agent’s reach. Runs on protocore-exp, an experimental branch of the Protocore core.

Changes to its own code depend on the mode: on a server, a pull request you approve; on a computer, applied on restart after checks; self-development can be turned off. The model and the key are yours. It is an open project for self-hosted use, not the Protocore enterprise platform. The film “The Task Board” (Corridor H8, 2:57) is on the Daedalus site and on YouTube.

Daedalus site · The film “The Task Board” · Source code

swiftclf-tuna — bilingual intent classification

Hierarchical intent classification (RU/EN) for routing AI assistants: retrieval, generation, tools, agent, clarification or fallback. Done as AI-orchestrated research with human-in-the-loop validation; CPU-first inference. Methodology and metrics are in the engineering notes.

github.com/ascorblack/swiftclf-tuna-research