Document

Resume

Alexander Stanislavovich Tikhonov

AI Systems Engineer · LLM Agent Architect

Astana, Kazakhstan GitHub: github.com/ascorblack Telegram: @notsoulmate Email: ascorblack@gmail.com · a@scorblack.ru Protocore: protocore.ascorblack.ru

Professional profile

AI Systems Engineer / LLM Agent Architect with two and a half years of commercial production experience in LLM/RAG systems, backend infrastructure for AI, semantic search, recommendation systems, local LLM inference, and agent orchestration.

Since January 9, 2024, I have worked at EBS Lan LLC (Russian academic e-library platform) under an official employment contract (contract title — ML engineer; actual role — AI Systems Engineer / LLM Agent Architect). I own the backend and infrastructure implementation of the company's key AI services: architecture, development, integration, deployment, monitoring, and production support of educational AI products.

My core strength is engineering production systems around LLMs, not generic model training: service design, RAG/backend architecture, integrations, resilience, observability, and operating AI in real infrastructure under strict latency and cost constraints.

Work experience

EBS Lan LLC

ML engineer (per contract) · AI Systems Engineer / LLM Agent Architect (actual role) January 2024 — present

I own the backend and infrastructure implementation of the company's key AI services: architecture, development, integration, deployment, monitoring, and production support.

LanGPT — production RAG/LLM assistant for academic tasks

  • Built the backend side of the product: the public API, the RAG service, the LLM backend balancer, and the administration components; I take part in their deployment and support in Kubernetes.
  • Public API with JWT authentication, rate limiting by pricing plan, and SSE streaming of responses; generation tasks run through RabbitMQ.
  • RAG service over Elasticsearch for working with verified educational and academic content from EBS Lan, available to the company through partner agreements with publishers, without reaching out to the external internet.
  • LLM backend balancer with a dynamic server registry, least-loaded / least-connections selection with priority, fallback, and OpenAI-compatible adapters.
  • The AI processing pipeline covers more than ten academic-writing tasks: hypothesis formulation, summarization, drafting sections with citations, content checking, and others; it includes content moderation of user requests.
  • Production monitoring on Prometheus, Grafana, and Elasticsearch.
  • The service is being piloted at 20 Russian universities; per the company's public materials, more than 37,000 requests and 6,500 active users during the testing period. The role and area of responsibility can be confirmed by the employer on official request.

AI-Snippets — generating search snippets over EBS Lan results

  • Pipeline: check the two-level cache (exact + semantic), classify the query, expand the query, multi-source search over Elasticsearch, assemble context (BM25 + RRF), and generate on vLLM with a final quality gate.
  • Polling scheme via Redis, circuit breaker, and graceful degradation: on any problem the snippet simply is not shown, with no error surfaced.
  • Stress test on the target GPU hardware: 13.2 hours, 55,200 requests, zero infrastructure errors (the service did not crash or return 5xx); quality gate OK rate of 87.3%; median client-side latency of 4.6 seconds against a 5 s SLA.
  • Serious tuning of vLLM parameters for snippet performance on a specific server configuration.

Books Keywords Generator — automatic catalog enrichment

  • Two-stage LLM generation of keywords for EBS Lan books based on the table of contents — first generation, then validation by a separate prompt with checks on language, length, uniqueness, and meaningfulness.
  • Dynamic context-length management, parallel processing with GPU overload protection.
  • Structured JSON logs in Elasticsearch for centralized monitoring of generation quality.

Search 2.0 — EBS Lan search platform

  • A platform of several production services: the main search API, the subject-area and author classifier, a vectorizer with caching, a spell checker, and search over educational video.
  • Hybrid BM25 and vector search over Elasticsearch, with a custom coefficient system and boosts by relevance and year of publication.
  • All ML components are designed CPU-only — no GPU dependency and with strict resource limits.

Independent projects

Protocore — protocore.ascorblack.ru

A self-hosted / on-premise multi-agent platform for organizations with data sovereignty requirements and their own AI infrastructure. Role — founder / solo developer: I designed and built the core runtime, enterprise backend, sandbox, chat and admin frontend applications, and the Helm / Kubernetes deployment.

  • An independent product under the ascorblack-labs organization; LanAgent for EBS Lan is the first integration track under a licensing model.
  • Protocol-first architecture: the orchestration runtime extends through typed contracts, hooks, and execution strategies; subagent orchestration and context compression for long sessions.
  • Enterprise layer: REST API with SSE streaming, RBAC, access plans with quotas and audit, sandboxed tool execution (gVisor), observability, and an autonomous worker for background tasks.
  • Stack — FastAPI, Next.js 15, PostgreSQL, Valkey (Redis), RabbitMQ, OpenSearch, Helm / Kubernetes; all infrastructure is self-owned: GitLab, Harbor, k3s.
  • Maturity: 7,500+ tests across the core and enterprise packages, 96.01% coverage in core, 84.07% in enterprise; ruff / mypy / bandit / pip-audit clean; beta stage.

swiftclf-tuna — github.com/ascorblack/swiftclf-tuna-research

An AI-orchestrated research artifact: bilingual (RU/EN) hierarchical intent classification for routing AI assistants. My role is problem framing, taxonomy and evaluation-criteria design, directing the agentic workflow, and human-in-the-loop validation of results.

  • Taxonomy of 4 L1 / 22 L2 classes; a reproducible holdout set of 2,625 examples; the metrics are used as a verifiable applied result, not a SOTA claim.
  • Selective prediction with a fallback branch and risk-aware metrics; probability calibration down to ECE 0.061; coverage @ risk ≤ 0.14 — 87.24%; accepted accuracy 87.07%.
  • CPU-first deployment: 278M parameters, p95 77 ms on a single CPU core; artifacts include model cards, training / eval pipelines, and holdout reports.

Open source: tools for AI agents

  • ai-audit-kit — a portable .audit standard: a reproducible AI code-audit pipeline (audit → triage → fix → review).
  • notify-telegram-cli — Telegram notifications from autonomous agents; pure standard library, zero dependencies.
  • kb-genesis — bootstrapping .kb/ knowledge bases for agent environments.
  • LightUniLLM — a lightweight framework for LLMs over the OpenAI Responses API: Pydantic, structured outputs, streaming.

Technology stack

LLM & AI Systems

  • LLM services, RAG, agent orchestration
  • Local LLM inference, vLLM tuning (max-num-seqs, KV-cache optimization, prefix-aware prompt design, speculative decoding experiments)
  • OpenAI-compatible adapters, dynamic LLM provider registry
  • Prompt & runtime design, tool routing
  • Selective prediction, calibration, quality gates (LLM-based content moderation, logprobs-based confidence scoring)

Backend

  • Python, FastAPI, async services
  • SSE streaming, JWT, REST API
  • Async API patterns: task queues with prioritization, SSE streaming with session recovery (Redis Pub/Sub + TTL cache), graceful degradation
  • Microservices, background workers, stream processing, pipeline architecture

Data & Search

  • Elasticsearch: BM25, dense vector search (kNN), hybrid combinations with RRF
  • Semantic reranking, RRF fusion
  • sentence-transformers, multilingual encoders
  • Recommendation systems, semantic profiling

Infrastructure & Ops

  • Docker, Nginx, Kubernetes / Helm deployment, gVisor
  • Redis (pub/sub, cache, rate limit), RabbitMQ
  • PostgreSQL, MariaDB / MySQL
  • Prometheus, Grafana, structured logging, OTLP
  • Self-hosted deployment: GitLab CI, Harbor, k3s
  • Latency- and cost-aware design, observability and debuggability in production

AI-Orchestrated Research

  • Problem framing, methodology design, and result validation for autonomous AI-agent workflows
  • Human-in-the-loop review, validation, counter-review
  • Model cards, holdout evaluations, risk-aware reports
  • Negative results as part of the research process

Languages

  • Russian — native
  • English — technical reading / listening; written and spoken communication — developing
← Back home