Temirlan Dzhoroevbuilds AI that ships.
I own the production agent orchestration layer at Redbrick — tool-calling agents with checkpointed state and memory across 40+ tenants and ~30K requests a day — plus the hybrid RAG beneath it and the eval and moderation stack that decides what ships. Six case studies below, each with its architecture, tradeoffs, and numbers, including the creation platform I scaled to 10M+ sign-ups.
- Open to AI engineering roles
- 12 years in Korea
- Fluent Korean
- Korean workplace experience
- No visa sponsorship required
Selected
work
Six systems shipped at Redbrick — hybrid RAG with reranking, agent orchestration, streaming codegen, LLM serving, and a platform that reached ten million users. Open any card for the full case study.
Enterprise Agent Platform
Hybrid parsing, hybrid retrieval, cross-encoder reranking — and an agent per tenant on top.
Companion Chat Runtime
A real-time AI companion that remembers, reacts in its own voice, and moves the scene.
Prompt to Playable
Type an idea, watch a playable game build itself — in the browser, mid-stream.
PDF to Lesson
A textbook PDF becomes an interactive lesson — generated multimodally, moderated for the room it runs in.
Model Gateway
The serving layer under both products — routing, moderation, caching, and back-pressure.
3D Studio
A browser 3D game studio that reached 10M users — block scripting, authoritative multiplayer, in-editor AI.
A bit
about me.
AI engineer with an HCI research background and three years of production shipping at Redbrick, Seoul.
Research, engineering, and product held in a single head — so ideas ship as fast as they’re proven.
I take AI from prototype to production — the messy middle where demos meet real users, latency, cost, and evals. Three years of it at Redbrick: the agent orchestration layer for 40+ tenants, retrieval that actually recalls, and the eval and moderation stack that decides what ships.
Before that, a real-time 3D engine behind 54M+ game plays and two years of HCI research — ten publications, six patents. I care about models that are correct, fast, and honest about their limits. The best AI feels less like magic and more like a tool you can actually trust.
Twelve years in Korea, fluent Korean, and a Korean workplace behind me. I work in either language, and I don’t need visa sponsorship.
Agents & Multi-Agent Systems
LangGraph runtimes, tool registries, checkpointed state and per-agent memory, human-in-the-loop escalation, and control flow you can reproduce.
RAG & Retrieval
Hybrid dense + lexical search, structure-aware ingestion, reranking, and ACL inside the query.
LLM Infrastructure
Gateways, key pools, fallback chains, semantic caching, and streaming that survives real load.
Evals & Moderation
LangSmith regression sets, multi-category classifiers with severity triage and escalation, PII redaction, and the observability to prove any of it.
Toolbox
The stack behind the case studies, grouped the way the CV lists it.
- Python
- TypeScript
- JavaScript
- C / C++
- SQL
- LangGraph
- LangChain
- LangSmith
- LiteLLM
- Vercel AI SDK
- OpenAI / Claude / Gemini APIs
- Multi-model routing
- ReAct tool loops
- Checkpointed state & memory
- Structured outputs
- Human-in-the-loop escalation
- Milvus
- pgvector
- ChromaDB
- LlamaIndex
- PostgreSQL
- Redis
- BM25 / hybrid search
- Cross-encoder rerankers
- Multi-category classifiers
- Severity triage & escalation
- PII redaction
- Eval datasets & regression harnesses
- FastAPI
- Express
- Kubernetes · Helm
- AWS S3 · ECR
- Prometheus
- Sentry
- E2B sandboxes
- React
- Next.js
- Three.js
- WebGL
- Vite
- Node.js
- Socket.IO
Where the
work happened
Three years of production AI at Redbrick, on top of an embedded systems and HCI research background. Same bullets as the CV.
AI Engineer — Redbrick, Seoul
Own the production agent orchestration layer, the retrieval under it, and the moderation and evaluation stack that gates every release.
- Run tool-calling agents on LangGraph for 40+ tenants at ~30K requests/day — checkpointed state and per-agent memory, up to 8 dependent tool calls per request; checkpoint recovery raised multi-step task success 82% → 96%.
- Secured a $1.2M government R&D grant with the end-to-end game-creation system: natural-language prompt to playable 3D game — intent parsing, asset generation, scene composition, logic synthesis.
- Cut hallucination rate 18% → 4% and lifted answer relevance +35% on versioned eval sets by shipping production RAG over 15K+ documents — hybrid dense + BM25 retrieval, query rewriting, cross-encoder reranking.
- Halved cost per answer (−48%, 21K → 12K tokens) and dropped latency 3.2s → 0.8s with rerank-then-budget context assembly and a semantic cache that removed ~40% of redundant LLM calls — about $2K/month saved.
- Built and own the moderation and evaluation stack: LangSmith datasets with 600+ regression cases, a 13-category classifier with 4 severity tiers that escalates to a human, fail-closed provider checks, and sub-10ms local guardrails.
3D Frontend Engineer — Redbrick, Seoul
Scaled the browser-based 3D engine behind a creation platform used by millions of young makers.
- Scaled a Three.js / WebGL / TypeScript game engine powering a platform with 10M+ sign-ups and 54M+ game plays; games I shipped on it passed 1M+ combined plays.
- Led the migration of a 50K-line codebase from Webpack to Vite with a rebuilt render pipeline and a replaced physics engine — build times −60%, bundles −35%, hot reload 8s → under 1s.
- Refactored the engine into modular, testable packages to unblock 3 teams working in parallel; added integration tests over the scene graph and asset loading.
- Redesigned onboarding from session-replay analysis and user interviews — 7-day activation +25%, Day-30 retention +15%.
Embedded Systems & HCI Researcher — DECS Lab, UNIST, Ulsan
Firmware and study infrastructure for a human–computer interaction lab.
- Wrote real-time IoT firmware in C for ARM Cortex-M MCUs driving 50+ networked sensor devices at sub-millisecond latency — the data pipeline behind 6 published studies.
- 6 conference papers, 4 journal articles, and 6 co-filed patents; first author at IEEE RO-MAN 2023, ICROS 2022, and HCI Korea 2022.
Education
M.S. in Design, Human–Computer Interaction — UNIST
GPA 4.0 / 4.3 · Lotte Scholarship. Thesis: human perception of social robot face and colour expression using computational emotion models.
B.S. in Computer Science & Industrial Design — UNIST
Global UNISTAR Silver Scholarship.
Research
Six conference papers, four journal articles, and six co-filed patents from the DECS Lab years — first author at IEEE RO-MAN 2023, ICROS 2022, and HCI Korea 2022.
Recent
coursework
Agents, retrieval, and generative AI — newest first. Each row carries its credential ID; linked rows open the issuer’s verification page.
Agentic AI
Advanced RAG with Vector Databases and Retrievers
Vector Databases for RAG: An Introduction
Foundation — Introduction to LangChain (Python)
Develop Generative AI Applications: Get Started
Build RAG Applications: Get Started
Gen AI Agents: Transform Your Organization
Hiring for agents, retrieval,
or LLM infrastructure?
Email is the fastest way to reach me — in English or Korean.
dzhoroev1@gmail.com