Skip to content
View olanokhin's full-sized avatar
🏠
Working from home
🏠
Working from home

Block or report olanokhin

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
olanokhin/README.md

Alex Anokhin

LLM Systems Engineer · Independent Researcher
Heilbronn, Germany (Remote-first, CET) · olanokhin@gmail.com · LinkedIn · olanokhin.com


Research

Project Description Status
Bitplane-ANN Memory-efficient progressive vector search: stores embeddings as compact 8-bit codes and filters candidates before reranking. Reached 99.8–99.9% Recall@10 on 1M-vector text and vision benchmarks with ~4× smaller storage than float32 Zenodo artifact · Paper
Agent Security Skill Portable OWASP-aligned security review skill for coding agents. Turns OWASP AI security guidance into active Claude Code/Codex behavior, covering 33 review categories across LLM, RAG, MCP, tool, and agentic systems Native Claude/Codex skills · Benchmark in progress
CPAR Cross-provider blind peer review — prevents single-model bias and hallucination. Outperformed single-model baseline 15/15 criteria in blind A/B evaluation (GLM-5 judge) Working demo
ATA Protocol Cryptographic Authorization Type Attestation for internet protocols — the "A" layer for TLS, MLS, QUIC IETF Draft-00
TurboQuant Benchmark Hadamard + Lloyd-Max compression on GloVe-200; Recall@10 vs compression across 1–8 bits; 8-bit sweet spot: 3.1× smaller, −0.74pp recall HF Space · GitHub
RCI Recursive Convergent Inference — dynamic MoE with convergence-gated stopping. Unexpected finding: model-relative complexity diverges from human difficulty labels Paper + Colab ready

Engineering

AI & LLM: RAG · Hybrid RAG · Graph RAG · Autonomous Agents (ReAct, multi-agent) · MCP · Transformers · LLM/RAG Evaluation
Stack: Python · TypeScript/JS · C/C++ · Kotlin · FastAPI · Streamlit · Gradio · ReactJS
Infra: Docker · Linux · Azure · Azure AI Foundry · Git/CI-CD


Selected Work

  • 🏗️ Exyte — Reduced LLM calls ~8× (1M → ~115k rows) through context engineering. Built hybrid RAG system replacing manual SharePoint/Excel search for prequalification questionnaires.
  • 🔬 Wire Germany — Proved feasibility of AI note-taking in a single day, pitched to 100+ people including CTO — greenlit for production. Cut meeting review time ~4× as sole AI engineer in cross-functional implementation.
  • 🥇 3× Hackathon winner — Future City (HHN & 42), Make.com (42 Heilbronn), Wire × Schwarz IT
  • 🎓 42 Heilbronn — Founded AI Club (78 members); built and deployed Hybrid RAG FAQ agent on Notion wiki

Interests

Protocol design · Cognitive architectures · Real-time AI systems · Live music · Basketball


Building intelligence. Expanding humanity.


Let's talk

Open to senior/staff AI engineer roles at AI-native companies — remote, English-first.
olanokhin@gmail.com · LinkedIn

Pinned Loading

  1. cpar-framework cpar-framework Public

    Cross-provider blind peer review — prevents single-model bias and hallucination. Outperformed single-model baseline 15/15 criteria.

    Python 1

  2. ata-protocol ata-protocol Public

    Cryptographic nature verification for internet protocols — the "A" layer for TLS, MLS, QUIC

    Python 1

  3. rci-inference rci-inference Public

    Recursive Convergent Inference — dynamic MoE with convergence-gated stopping. Unexpected finding: model-relative complexity diverges from human difficulty labels

    Jupyter Notebook

  4. hnsw-turboquant-glove hnsw-turboquant-glove Public

    TurboQuant-style HNSW benchmark on GloVe-200: Recall@10 vs compression across 1-8 bits.

    Python 1