Dmitry Zorin
Dmitry Zorin@medonomator·Batumi, GE
Booking · Q3 2026

AI engineer.
I build LLM systems that survive real users.

The plumbing under products with paying users: token-level billing, cognitive memory, multi-stage evals, voice.

Three years on AI infra. Twelve on backend.

2026-05 → MindForge v2 shipped: GPT-5 routing, cognitive memory, pgvector · mind-forge.org
Start a projectSee selected work
y
TypeScript
y
LLM systems in production
systems in production
01What I build

Four areas, each shipped to production more than once.

The cases that break under load - cost runaway, hallucinated entities, provider outages, prompt drift, rate-limit storms. I plan for them up front, since patching after the first incident takes ten times longer.

LLM infrastructure

Token-level billing across text, TTS, and Whisper. Provider fallback chains. Prompt versioning with hash-based drift detection. Full Langfuse traces from request to model output.

billingevalsobservability

Cognitive memory

Episodic and semantic memory over Qdrant or pgvector. Bayesian Knowledge Tracing for skill mastery. Hidden Markov Models for emotional state. Thompson sampling for adaptive content.

memoryBKTbandits

AI pipelines

Multi-stage quality gates with Zod-validated outputs. Entity reference checks against retrieved context. Retry loops that pass rejection reasons back into the prompt instead of blind regeneration.

RAGvalidationagents

Multi-modal

Streaming chat over μWebSockets. TTS and Whisper with hallucination sanitization on both ends. Function calling with project-aware context. Real-time interrupts.

voicestreamingWS
02Selected work

Four projects in production - one open, three under NDA.

Adaptive learning
Solo architect and engineer

MindForge. AI engineering curriculum.

64 lessons of AI engineering, taught by a Socratic tutor that grades exercises semantically and adapts coding tasks to the learner's GitHub history. Whole stack mine, from database to UI.

  • Prompt composition framework. 15+ task prompts, hash-based drift detection on every change
  • Token-level billing across text, TTS, Whisper, and function calls
  • WebSocket streaming via μWebSockets.js for chat, hints, and live transcription
  • GS2 lesson schema validated with Zod, with misconception flags attached per concept
Visit MindForge
Behavior change
Lead AI engineer, NDA

Cognitive notification engine for a habit app.

Per-user push generation for ~2k MAU. A Thompson-sampling bandit learns the tone and angle each user responds to. Four-stage LLM quality gate: generate, judge, entity-check against context, fallback. When the HMM detects emotional risk, an override forces the model into supportive tone.

  • 8 notification types with tier-based budgets (ACTIVE, SLOWING, DORMANT, COLD)
  • Bayesian Knowledge Tracing for habit mastery probabilities
  • Hidden Markov Model for emotional trajectory and crisis prediction
  • Adaptive timing on a Redis bloom prefilter with SQL verify pass on hits
Long-context memory
Architect and engineer

Episodic and semantic memory layer.

A cognitive context service that enriches every LLM prompt with relevant past interactions, learned facts as subject-predicate-object triples, emotional trends, and behavioral predictions. Each module degrades gracefully on its own.

  • OpenAI text-embedding-3-small over Qdrant with filtered similarity
  • Survival analysis for churn prediction, injected as warnings into the prompt
  • Concept ingestion from PR reviews feeds automatic mastery tracking
  • Sub-100ms P95 enrichment with async-safe failure modes
Developer tools
Solo build

PR-aware AI tutor for production repos.

A GitHub App that generates personalized coding tasks against the learner's own repos, then reviews their PRs with inline comments. Difficulty calibrates from past review decisions: three rounds of changes make the next task easier, two clean approvals raise the bar.

  • Octokit-based diff fetching with line-number validation and SHA-dedup
  • Stack-aware persona for ML, distributed systems, or web, each with its own antipatterns
  • Concept extraction from PR review flows into the learner mastery model
  • BullMQ async pipeline, observability via Sentry and Langfuse
03Shipping log

What actually went out - last weeks, dated, no embellishment.

A trimmed slice of recent commits across mind-forge, exocortex, confyday-back, second-brain, math-research. Maintenance commits, lint passes, and merges removed.

  1. 2026-05-09
    mind-forge
    voice mentor vocab hints + billing energy reset countdown
  2. 2026-05-08
    mind-forge
    voice mentor multi-layer orchestration (intent + facts + planner + critic)
  3. 2026-05-07
    mind-forge
    native streaming for gpt-5 reasoning models, per-chunk TTS
  4. 2026-05-06
    mind-forge
    energy billing rewrite + voice mentor observability
  5. 2026-05-06
    mind-forge
    pre-flight model gate + estimate-shortfall clamp
  6. 2026-05-05
    mind-forge
    PWA last-route persistence + Lab dashboard expansion
  7. 2026-05-04
    mind-forge
    ai-engineering course +10 lessons (59-68), full candy audit pass on 68
  8. 2026-05-04
    mind-forge
    project-lab v2: portfolio signing, leaderboard, PR review prompt v2
  9. 2026-05-04
    ai-pulse
    weekly LLM industry pulse via Claude CLI to Telegram, sources-first strategy
  10. 2026-05-03
    confyday-back
    per-locale push source/keywords (CON-847)
  11. 2026-05-02
    confyday-back
    curated push library replaces LLM/fact-pushes pipeline
  12. 2026-05-01
    mind-forge
    live-practice pendingCorrections + selective corrections + fluency turns
  13. 2026-05-01
    mind-forge
    learning-theory lessons lt-11 margin-bounds, lt-12 online-regret, lt-13 deep-generalization
  14. 2026-05-01
    mind-forge
    CS lessons 03-04 candy rewrite (51 + 53 topics RU+EN)
  15. 2026-04-30
    mind-forge
    mathlikeanim block type + consistent-hashing animation pilot
  16. 2026-04-30
    mind-forge
    CS lesson 02 candy rewrite, all 50 topics RU+EN
  17. 2026-04-29
    mind-forge
    lessons hot-reload + schema-driven cross-link filter
  18. 2026-04-29
    mind-forge
    forbidden-phrase ratchet gate + LaTeX validator wired into CI
  19. 2026-04-28
    mind-forge
    Yandex sign-in (RU), web-push permissions, unified PWA manifest
  20. 2026-04-27
    math-research
    optimal hallucination scorer - AUROC 0.979 QA, 0.772 Sum (pure gzip)
  21. 2026-04-26
    math-research
    Kolmogorov Structure Function + 4-method benchmark
  22. 2026-04-26
    math-research
    hallucination detection benchmark on HaluEval
  23. 2026-04-25
    exocortex
    cross-domain noise gates + Unicode-aware skip patterns in recall
  24. 2026-04-24
    exocortex
    HyDE for short-query embedding lift + LLM-as-reranker (gpt-4o-mini listwise)
  25. 2026-04-24
    exocortex
    graph-first retrieval via mem_item_entities
  26. 2026-04-24
    second-brain
    astrology weekly transit alerts (EN Pro feature)
  27. 2026-04-24
    second-brain
    extract cities.ts (12k LOC) to JSON asset
  28. 2026-04-23
    exocortex
    cognitive memory v2 - universal items, lifecycle, multi-signal recall
  29. 2026-04-23
    second-brain
    migrate runtime and CI from PM2 to Docker
  30. 2026-04-21
    confyday-back
    langfuse prompt management - centralized constants, datasets, experiment runner
  31. 2026-04-21
    confyday-back
    LLM quality gate + 7-day cross-type anti-repetition for pushes
  32. 2026-04-19
    confyday-back
    BKT/survival models bug - every habit tracking treated as failure
  33. 2026-04-19
    second-brain
    EN i18n + Stripe payments for Astro AI bot, locale-aware system prompts
  34. 2026-04-18
    exocortex
    cheap swarm testing - replay-consensus and --cheap mode
  35. 2026-04-17
    exocortex
    Langfuse swarm tracer - 20 event types as spans/generations/events
  36. 2026-04-17
    exocortex
    evolution engine v2 - swarm-evolution adapter, enhanced GA
  37. 2026-04-17
    confyday-back
    rolling deploy for zero-downtime production updates
  38. 2026-04-17
    confyday-back
    ban em dashes and AI-typical patterns from push notifications
  39. 2026-04-16
    exocortex
    file-triggered intelligence - per-file context on Read/Edit
  40. 2026-04-16
    exocortex
    git history indexing - batch indexer, commit hook, recall integration
  41. 2026-04-16
    exocortex
    PostgreSQL graph, GA self-optimization, memory consolidation
  42. 2026-04-15
    confyday-back
    gpt-5.4-mini upgrade, sampling params for premium users
  43. 2026-04-15
    confyday-back
    living mind - narrative threads, change detection, follow-up awareness
  44. 2026-04-15
    confyday-back
    ai_notification push support - runtime type guards, entityId deep linking
  45. 2026-02-06
    second-brain
    YooKassa payment polling cron as webhook fallback
  46. 2025-12-07
    second-brain
    achievements system + brain balance economy
  47. 2025-12-06
    second-brain
    brain balance system, living mind, streaming responses

47 entries · pulled from public + private repos · last refreshed 2026-05-09

04What clients say

Two clients, both with AI live in production today.

“Dmitry Zorin shipped our entire AI layer - voice mode, user-portrait personalization, structured-output notifications, cost-tier model routing. Eval coverage from day one, Langfuse observability baked in. The kind of engineer you keep on speed dial.”

AP
Alex PikunovFounder · Confyday

“Built our LLM evaluation harness from scratch. Structured outputs, prompt-judging for candidate scoring, regression tests on every model swap. We stopped deploying AI on vibes - every prompt change now has a measurable delta before it ships.”

FL
FlomniAI engineering team
05Stack

What I reach for. Boring infra, sharp on the AI layer.

TypeScript across the stack, Python at the AI layer. Postgres, Docker, and Hetzner for plumbing. Specialized tools where the difference shows up in the bill.

Models
gpt-5 (default reasoning) · gpt-5.4 · gpt-5.4-mini (premium routing) · gpt-4o · gpt-4o-mini (rerankers, judges) · o1 · o1-mini · Claude Opus · Claude Sonnet · Claude Haiku · text-embedding-3-small · Whisper-1 · TTS-1
LLM tooling
Langfuse (traces + prompt mgmt) · Zod schemas · JSON Schema · structured outputs · function calling · HyDE · listwise rerank · Thompson sampling · BKT · HMM · ts-fsrs (spaced repetition) · Promptfoo
Vector + queue
Qdrant (most projects) · pgvector (Postgres-only stacks) · Redis bloom · KeyDB (ARM, not :alpine) · BullMQ · Postgres 16 · HNSW
Backend
TypeScript (default) · Python (AI layer only) · NestJS · Express · FastAPI · TypeORM · μWebSockets.js · Pino · Sentry · Stripe · YooKassa
Frontend
Next.js 16 · React 19 · Tailwind v4 · Zustand · TanStack Query · Framer Motion · PWA + service worker · mathlikeanim-rs
Infra
Docker (linux/amd64 builds from M1) · Hetzner (most prod) · Oracle Cloud (free-tier ARM) · GCP (fallback) · DigitalOcean · Cloudflare · nginx · GitHub Actions · GitLab CI · PM2 · systemd · certbot (Let's Encrypt)
06How to engage

Three engagement modes - all of them end with code in your repo.

Fixed scope, fixed price. I bill outcomes instead of hours, so every engagement leaves you with code and docs you can actually hand to the next engineer.

4 to 12 weeksfrom $12K
Build

One AI feature, spec to production.

I take a defined system from spec to deploy: cognitive memory, RAG, evals harness, voice mode, or a billing layer. Weekly demos against a checklist you signed off on, no scope creep.

  • Architecture doc and delivery plan up front
  • Deployed to your infra, not mine
  • Evals and observability shipped with the feature
Discuss this engagement
1 to 2 weeksfrom $4K
Audit

Architecture review of a live system.

Your AI system is in production but something feels off - costs creeping, hallucinations slipping past your checks, latency spiking on hot paths. I find what is broken and write a prioritized fix plan with effort estimates against it.

  • Code and infra walkthrough
  • Cost, risk, and drift report
  • Remediation roadmap with effort estimates
Discuss this engagement
monthly engagementfrom $5K/month
Embed

Senior engineer on your team.

I drop into your team for a focused push. Pair on the hardest features, ship them, leave behind PRs and design docs your engineers can extend. No onboarding ramp on basics.

  • Delivery measured in merged PRs
  • Architecture pairing with your leads
  • Decisions captured in writing
Discuss this engagement