AI Systems Engineer LLM Agent Architect Astana · Kazakhstan

Aleksandr Tikhonov.

I design production systems around LLMs: RAG services, agent orchestration, semantic search, and self-hosted inference — under strict latency and cost constraints.

2.5+
years of production AI
20
universities in pilot
11,000+
Protocore tests

Reference architecture

FIG. 1 · GENERAL VIEW
chat / web clientsSSE streaming · RU/ENin prod: LanGPT · Protocore Chat
admin dashboardcontrol plane · RBACin prod: Protocore Dashboard
API · agent orchestrationFastAPI · contracts · quality gatesin prod: LanGPT API · Protocore Runtime
searchBM25 + kNN + RRFin prod: Search 2.0 · AI-Snippets
LLM inferencevLLM · load balancerin prod: LanGPT · AI-Snippets
sandboxgVisor · k8s podsin prod: Protocore Sandbox
infrastructureKubernetes · Helm · Prometheus · RabbitMQ · Redis · Postgresin prod: EBS Lan · Protocore
THE STACK I WORK WITH, DRAWN AS A SYSTEM · HOVER A NODE — WHERE IT RUNS IN PROD

A diagram beats a bullet list. This exact scheme runs end-to-end in my self-hosted platform Protocore.

Flagship SYS—01 OPEN CORE
Lead project · ascorblack-labs · open core, MPL-2.0

Protocore.

A protocol-first runtime for stateful, tool-using LLM agents. The core — "the agent loop, and nothing else" — is open under MPL-2.0 and installs via pip; the platform on top of it (enterprise backend, sandbox, chat, dashboard) is proprietary and runs entirely on my own infrastructure. Flagship project — founder / solo.

  • Open core · MPL-2.0 · pip install protocore
  • 20 Protocol interfaces · zero upward imports
  • Snapshot / resume of runs · context compaction
  • 11,000+ tests · reproducible research
  • Enterprise layer: RBAC · gVisor sandbox · SSE

Systems index

8 entries · 2024—2026
SYS—02
PROD

LanGPT

RAG/LLM assistant for academic tasks on the EBS Lan (Russian academic e-library platform). Public API with JWT and SSE, RAG over Elasticsearch, LLM backend balancer, moderation.

20 universities in pilot
37,000+ requests
6,500 users
SYS—03
PROD

AI-Snippets

Snippet answers over EBS Lan search results: cache, BM25+RRF, quality gate, circuit breaker, graceful degradation. Stress test: 13.2 hours without a single infrastructure error.

55,200 requests · 0 × 5xx
median 4.6 s / SLA 5 s
SYS—04
PROD

Search 2.0

EBS Lan search platform: hybrid BM25 and vector search, subject-area classifier, spell checker, video search. All ML components are CPU-only.

hybrid BM25 + kNN + RRF
ML: CPU-only
SYS—05
ACTIVE

Daedalus

A personal self-developing agent on the Protocore core: it lives in Telegram and in its own web app, runs inside a Linux container with real tools, and changes its own code through PRs approved from the chat.

Telegram + PWA · self-development via PRs
supervisor · snapshots · weeks of uptime
MIT · 320+ tests
SYS—06
R&D

swiftclf-tuna

Bilingual (RU/EN) hierarchical intent classification for routing AI assistants. AI-orchestrated research: selective prediction, calibration, CPU-first deployment.

accuracy 87.07% · ECE 0.061
p95 77 ms / 1 CPU
SYS—07
OSS

AI Audit Kit

A portable .audit standard: turns AI code audit, triage, and bulk fixes into a reproducible pipeline with artifacts instead of chat sessions.

audit → triage → fix → review
open source
SYS—08
OSS

Agent Tooling

Tools for autonomous agents: notify-telegram-cli — dependency-free notifications, kb-genesis — knowledge bases for agent environments, ai-notes, LightUniLLM — a lightweight LLM framework.

notify-cli · kb-genesis · LightUniLLM
github.com/ascorblack

Profile

Since January 2024 — AI Systems Engineer at EBS Lan (Russian academic e-library platform): I own the backend and infrastructure side of the company's key AI services, from architecture to production support in Kubernetes.

In parallel — founder of ascorblack-labs: Protocore, a protocol-first runtime for LLM agents. The core is open under MPL-2.0; the platform is proprietary and runs entirely on my own infrastructure (GitLab, Harbor, k3s).

My strength is engineering production systems around LLMs rather than generic model training: reliability, observability, and degradation that never surfaces as an error.

Specification

  • LLM / RAG / agent orchestrationcore
  • Protocol-first agent runtime · context compactioncore
  • Python · FastAPI · asynciocore
  • Elasticsearch · BM25 + kNN + RRFcore
  • vLLM · local inferencecore
  • Kubernetes · Helm · gVisorops
  • RabbitMQ · Redis · PostgreSQLdata
  • Prometheus · Grafana · OTLPobs
  • Next.js · React — platform frontendsui