I build production LLM systems that take manual work off revenue teams — and I do the analysis on top of the data they produce. Business rules, prompts, retrieval, model orchestration, Python/TypeScript services, evaluation, CRM/BI delivery, unit economics — and then the report a decision is actually made on.

Portfolio · Lead Chat · Email

Five years in engineering, two in AI. Open to Senior / Lead Applied AI roles with end-to-end ownership. Remote, UTC+3, open to teams abroad.

Denominators for every number above: 94spec.github.io/#measure.

Quality is controlled with four task-specific golden sets, human calibration, schema validation, a typed error taxonomy, confidence intervals and regression gates. A frozen, manually adjudicated 500-deal golden set recorded 98% model-vs-expert decision agreement; critical-violation recall is gated separately, because an average score hides exactly the misses that matter.

For Lead Chat I own the product concept, AI logic and core architecture. A development partner works with me on the technical side: selected components and code testing.

The pipeline gives me structured data for every conversation. The second half of the job is turning that into answers the business acts on — studies of up to 3,000 deals, horizons up to six months, covering every employee who talks to customers:

• Why conversion dropped over a quarter — every substantive conversation for a product line across three months (empty calls excluded from the base, not diluted into percentages), dozens of features per deal, month over month. Observations kept separate from hypotheses.
• "Too expensive" is not always about price — a two-level reason taxonomy, with records that failed processing excluded explicitly rather than quietly. The top level says "cost"; the second level says the customer cannot get instalment financing. Different problem, different decision.
• Part of the sales team's load is not sales — inbound requests classified by type and by owner (sales, refunds, support), surfacing deals closed under the wrong reason and a conversion denominator that was measuring the wrong thing.

Rules I keep: sample control before any percentage, a denominator for every share, hypotheses labelled as hypotheses, every category backed by a customer quote, and an explicit limitations section stating what the data cannot show.

Business process → AI behaviour spec → architecture → implementation → evaluation → production integration → cost/quality optimisation → the report behind the decision

• Conversation intelligence and rubric-based AI quality control
• LLM evaluation: golden sets, LLM-as-a-Judge, error taxonomies, regression gates
• RAG: chunking, hybrid retrieval, grounding, source-aware answers
• AI agents: structured outputs, tool calling, memory, guardrails
• Voice AI and dialogue-state design
• Sales, marketing, product and pricing analytics on conversation data
• CRM/BI/API integration and multi-provider cost optimisation

• llm-evaluation-harness — dependency-free evaluation framework for four task families, with methodology, error taxonomy and a sample report.
• prompt-contracts — prompts as versioned contracts: strict output schema, a linter where every rule names the production failure it prevents, and a release-impact diff that says whether a change forces the golden set to be re-run.
• golden-set-builder — building the reference set itself: stratified sampling that reports what the pool could not give, annotator agreement before any model metric, adjudication that refuses majority vote, tamper-evident freezing and drift detection.
• voice-agent-runtime — the parts of a voice agent that are not the model: dialogue policy as a transition table, barge-in that never resumes, a per-stage latency budget judged at p95, line capacity, and guardrails that decide whether a reply can be spoken at all.
• rag-grounding-lab — what chunking costs and where each retrieval strategy loses: BM25 against Russian morphology, character n-grams against near-identical codes, rank fusion carrying a wrong result. Plus sentence-level grounding that separates an unsupported claim from a correct refusal.
• conversation-intelligence-lab — the analytics methodology above on synthetic data: strict output contract, traceable aggregates, architecture notes.
• 94spec.github.io — the portfolio itself: static, no build step, with CI that checks metadata, anchors and assets.

Both libraries run offline, with no API key and no third-party runtime dependencies.

In production, under my responsibility: Python TypeScript FastAPI Node.js PostgreSQL pgvector SQL Docker OpenAI API Anthropic API GigaChat YandexGPT Dify RAG AI Agents Structured Outputs Tool Calling LLM-as-a-Judge Voice AI CRM / BI Git

Worked with outside that loop: LangChain LangGraph Qdrant vLLM Hugging Face Fine-tuning / LoRA RAGAS A/B Testing Guardrails Prompt Injection NLP

RU: Строю production LLM-системы для коммерческих процессов и делаю на их данных аналитику, по которой принимают решения — по продажам, маркетингу, продукту и тарифам. Открыт к Senior / Lead Applied AI ролям, удалённо, UTC+3.