// olanokhin.com

Alex
Anokhin

LLM Systems Engineer · Independent Researcher
Building production-grade RAG pipelines, autonomous agents,
and infrastructure for the AI era.

Currently @ Exyte · Heilbronn, Germany
olanokhin@gmail.com linkedin.com/in/olanokhin github.com/olanokhin
57%
LLM cost reduction
@ Exyte
4×
Speech pipeline
speedup @ Wire
3×
SMM efficiency
gain @ ClyrAI

// about

I'm Alex — originally from Ukraine, now based in Germany. Before writing a line of AI code I spent 12 years as a live sound engineer — running up to 32 channel consoles at concerts, TV broadcasts, corporate events, and ceremonies. A live show is the only deploy you can't roll back. That shaped how I think about reliability.

The move to AI wasn't a pivot, it was pattern recognition. When I first saw autonomous agent architectures, I recognized the topology: parallel independent processing chains, signal routing, fail-states with no retry button. Same board — different domain.

Right now I'm an AI Engineer at Exyte, building document intelligence pipelines for construction data at scale — multilingual CSVs, OCR on noisy engineering documents, hybrid RAG on Azure. Outside of work I'm doing independent research on agent architectures, specifically how agents should reason about failure, not just succeed.

I'm married to Angelika, we have a 1-year-old daughter Anna. When I'm not shipping, I'm at a live gig somewhere, playing basketball, or at a table tennis table.

// work

2026 — present
AI Engineer
Exyte Management GmbH · Stuttgart, Germany
Enterprise
  • Hybrid RAG with graph retrieval on Azure for construction documentation — POs, multilingual CSVs with 1M+ rows across 20+ languages.
  • Reduced translation candidates from 1M → ~115k rows via layered language detection and regex filtering. 57% cost reduction vs naive full-pass approach.
  • Built HITL dashboard so non-technical stakeholders can manage glossary and monitor token-level costs in real time.
  • Evaluated Mistral Document-AI, MinerU, Docling, Doc.OCR for OCR on noisy engineering documents.
2025
AI Integration Engineer
Wire Germany GmbH · Berlin (Remote) · Apr – Sep 2025
Enterprise
  • Built a PoC in one evening — Whisper + LLM orchestration, fully local via Docker Compose, e2e encrypted. Demoed live to 100 people the next morning at sprint planning. The room went silent.
  • Microservice adopted into Wire's product roadmap — 75% reduction in meeting review time.
  • Kotlin/Whisper-JNI transcription service with concurrent audio stream processing — 4× speedup vs sequential baseline.
  • NLP-to-cron reminder bot with natural language scheduling and Markdown UI, integrated directly into Wire chat.
KIWI Knowhow
Co-founder & Tech Lead · Virtual Senior Engineer RAG Platform
Venture
  • 🥇 1st place + €10,000 AWS credits at CyberValley AI Founder Program.
  • RAG platform for instant technical retrieval — a "virtual senior engineer" on demand for engineering teams.
2024
ClyrAI
Freelance · AI Content Strategy Agent
Freelance
  • AI agent automating LinkedIn content strategy — 3× efficiency gain for SMM managers.
  • Custom Python parser cut data acquisition costs by 90%.
Abbey Tholey — Virtual Monk
Freelance · Context-Aware RAG Agent
Freelance
  • Digitized museum archives dating back to 634 AD — nearly 1,400 years of history — into an interactive RAG assistant with persistent session memory.
GreeterAI
Co-founder & Solo Developer · Generative Video Greetings
Venture
  • End-to-end generative AI video pipeline: LLM + xTTS + Wav2Lip for personalized greetings.
  • 300+ waitlist signups on demo day.
Tsunami 2025
Freelance · Career Intelligence Agent
Freelance
  • Conversational agent that maps user domains to labor market vulnerability quadrants and generates personalized 3-year upskilling roadmaps.

// research

BA² Framework — Anokhin²
Biomimetic recursive agent architecture for reliable long-horizon execution. Every node is identical: decompose until deterministically solvable, then crystallize as a validated skill. Graph-based memory eliminates context loss. Dual-channel failure system — Grit (budget) and Frustration (escalation with evidence) — replaces silent loops with structured reasoning about dead ends. Named after P.K. Anokhin, Soviet neuroscientist of Functional Systems Theory.
PoC · Paper in progress
ANV Protocol
Cryptographic nature verification for internet protocols — an "A" layer sitting alongside TLS, MLS, QUIC. Addresses the provability of human vs. automated origin for internet communications.
IETF Draft-00
RCI — Recursive Convergent Inference
Dynamic mixture-of-experts with convergence-gated stopping criteria. Inference halts when output confidence converges rather than at a fixed depth — trading compute for quality adaptively.
Paper + Colab ready
CPAR — Collaborative Peer AI Review
N independent AI reviewers with distinct cognitive profiles conducting blind iterative peer review until consensus. The paper itself was written using this method — 14 iterations with Claude as author and Grok / Gemini / ChatGPT as blind reviewers. All on free tiers, in one evening.
arXiv prep

// hackathons — 3× 1st place

🥇
Future City Hackathon — HHN & 42 Heilbronn
AI agent for automated grid connection requests — end-to-end processing of PV and energy storage infrastructure documents.
November 2025
🥇
Make.com Hackathon — 42 Heilbronn
Automated Next.js internship platform integrating 42 Intra API, Google Sheets, and email workflows.
July 2025
🥇
Wire × Schwarz IT Hackathon
Whisper voice transcription + chat summary AI features. Built the evening before the demo. Directly resulted in an internship offer at Wire.
November 2024

// writing & insights

32 channels = architectural thinking
Systems
"People see a board with 1000 knobs and say 'no thanks'. I see 32 identical strips. Same thing that happens when you look at a codebase."
Zero-rollback engineering
Reliability
"Kubernetes goes down — you restart it. The president's microphone drops mid-ceremony — you just died. That's where my convergence criteria come from."
Debugging hidden layers before I knew what they were called
Interpretability
"You can diagnose a broken stage monitor you can't physically hear by watching where the singer misses notes. Inferring hidden state from observable outputs. Literally ML interpretability by hand."
SaaS pricing model invented in a bar in Odessa
Product
"I invented tiered SaaS pricing + willingness-to-pay segmentation before I knew those terms existed. Y Combinator doesn't teach that one."
Writing a research paper with the method it describes
Research
"14 iterations. Claude as author. Grok, Gemini, ChatGPT as independent blind reviewers with web search. Stopped when reviewers said: go run the experiment. All on free tiers in one evening."
Load-balancing a threat across 700 nodes
Incident response
"The best incident response is when the system protects itself. I just had to say the right thing into the microphone."

// stack

Python TypeScript / JS C / C++ Kotlin RAG · Graph RAG Autonomous Agents MCP OCR · Document AI Whisper FastAPI ReactJS · Next.js Azure AI Foundry Docker · Linux Git / CI-CD