Aleksandr Tikhonov.
AI Systems Engineer · LLM Agent Architect
I design and operate production systems around LLMs: RAG services, agent orchestration, semantic search, and self-hosted inference — under strict latency and cost constraints.
01Profile
AI Systems Engineer / LLM Agent Architect with two and a half years of commercial production experience in LLM/RAG systems, AI backend infrastructure, semantic search, recommender systems, local LLM inference, and agent orchestration.
Since January 9, 2024 I have worked at EBS Lan LLC (a 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.
In parallel — founder of ascorblack-labs: Protocore, a protocol-first runtime for LLM agents. Its core is open under MPL-2.0 and installs via pip; the platform layer (enterprise backend, sandbox, chat, dashboard) is proprietary and runs entirely on my own infrastructure.
My strength is engineering production systems around LLMs rather than generic model training: service design, RAG/backend architecture, integrations, resilience, observability, and operating AI in real infrastructure under strict latency and cost constraints.
02Experience
EBS Lan LLC
ML engineer (per contract) · AI Systems Engineer / LLM Agent Architect (actual role)
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 maintenance 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 verified educational and academic content from the e-library, 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; includes content moderation of user requests.
- Production monitoring on Prometheus, Grafana, and Elasticsearch.
Figures come from the company's public materials. The role and area of responsibility can be confirmed by the employer on official request.
AI-Snippets — generating search snippets over EBS Lan search results
- Pipeline: two-level cache (exact + semantic), query classification and expansion, multi-source search over Elasticsearch, context assembly (BM25 + RRF), generation on vLLM with a final quality gate.
- Polling scheme via Redis, circuit breaker, and graceful degradation: on any problem the snippet is simply not shown, with no error surfaced.
- Tuning of vLLM parameters for snippet performance on a specific server configuration.
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 strict resource limits.
Books Keywords Generator — automatic catalog enrichment
- Two-stage LLM generation of keywords for e-library books from the table of contents — generation, then validation by a separate prompt checking language, length, uniqueness, and meaningfulness.
- Dynamic context-length management, parallel processing with GPU overload protection; structured JSON logs in Elasticsearch.
03Independent projects
Protocore — protocol-first runtime for LLM agents · founder / solo developer
protocore.ascorblack.com ↗The core is "the agent loop, and nothing else": the ReAct runtime, the context budget, the tool surface, compaction, and stop conditions. Everything outward-facing is a Protocol the host implements. Open under MPL-2.0 (ascorblack-labs/protocore-community, `pip install protocore`). The platform on top of the core is proprietary — enterprise backend, sandbox, chat, dashboard, autonomous worker, Helm/Kubernetes deployment — designed and built by me end to end.
- 20 Protocol interfaces (ILLMProvider, IRunStore, IToolRegistry, IMemory, IWorkspace…), zero upward imports — the core knows nothing about databases, HTTP, or deployment; direct/deep run strategies, delegation to subagents with width and depth budgets, snapshot and resume of runs.
- Three-layer tool surface (tenant policy → clipped surface → progressive discovery over BM25) and multi-tier context compaction; 500+ runtime constants in a frozen per-tenant snapshot.
- 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.
- A reproducible-research and publication site (protocore.ascorblack.com): methodology, limitations, data, and PDF papers.
An agent on the Protocore core that lives in Telegram and in its own web app, runs inside a Linux container with real tools, and changes its own code through pull requests approved right from the chat.
- One operator, one container: shell, browser, git, filesystem, long-running services, a scheduler, subagents, memory; each Telegram forum topic is a session with its own workspace.
- Built to run for weeks: sessions survive restarts, runs resume from snapshots, context is compacted instead of lost, spend is capped by a supervisor the agent cannot edit.
- React 19 web app (PWA): live transcripts, a file browser with previews, two sessions side by side; self-hosted SearXNG for web search, MCP servers per session.
swiftclf-tuna — AI-orchestrated research · bilingual intent classification
github.com/ascorblack/swiftclf-tuna-research ↗Hierarchical (RU/EN) intent classification for routing AI assistants. My role — 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 a verifiable applied result, not a SOTA claim.
- Selective prediction with a fallback branch and risk-aware metrics; probability calibration; CPU-first deployment with model cards, training / eval pipelines, and holdout reports.
Open source: tools for AI agents — grown out of daily work with autonomous agents
github.com/ascorblack ↗- 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 (Claude Code, Codex, Gemini CLI, Cursor).
- LightUniLLM — a lightweight LLM framework over the OpenAI Responses API: Pydantic, structured outputs, streaming.
04Technology stack
- LLM services, RAG, agent orchestration
- Local inference, vLLM tuning (max-num-seqs, KV cache, prefix-aware prompt design, speculative decoding)
- OpenAI-compatible adapters, dynamic provider registry
- Prompt & runtime design, tool routing, context compaction
- Selective prediction, calibration, quality gates, logprobs-based confidence
- Python, FastAPI, asyncio, Pydantic v2
- SSE streaming, JWT, REST APIs
- Prioritized task queues, SSE session recovery (Redis Pub/Sub + TTL cache), graceful degradation
- Microservices, background workers, pipeline architecture
- Elasticsearch / OpenSearch: BM25, kNN, hybrids with RRF
- Semantic reranking, sentence-transformers, multilingual encoders
- PostgreSQL + pgvector, MariaDB / MySQL, Redis / Valkey, RabbitMQ
- Recommender systems, semantic profiling
- Docker, Nginx, Kubernetes / Helm, k3s, gVisor
- Self-hosted: GitLab CI, Harbor, cert-manager
- Prometheus, Grafana, structured logging, OTLP
- Latency- and cost-aware design, observability and debuggability in production
- Next.js 15, React 19, TypeScript — platform frontends (chat, dashboard, PWA)
- Problem framing, methodology design, and result validation for autonomous AI-agent workflows
- Human-in-the-loop review, counter-review, reproducible reports
- Model cards, holdout evaluations, risk-aware reports, negative results as part of the process
05Languages
- Russian
- native
- English
- technical reading / listening; written and spoken communication — developing