Eduard Kharaev
Applied AI Engineer · Computer Vision · RAG · LLM Agents
I build production AI systems end to end: vision-language pipelines, retrieval with vector search, self-hosted LLM inference on GPUs, and agents with strict boundaries. Most recently Head of AI at a dairy FMCG group: I wrote the first production systems myself, then hired and led a team of seven engineers.
Writing
Engineering lessons from production systems.
Evidence, Not Answers: Why Our VLM Never Made the Final Decision
The vision-language model extracts evidence, deterministic rules decide, and "unknown" is a feature. Real audit numbers from a production shelf-recognition pipeline on self-hosted Qwen.
Read the article →Projects
Production systems I designed and built. Code is private; each case study covers architecture, decisions and measured results.
Retail Shelf Recognition
Production multimodal recognition of products on store shelves: GroundingDINO, Qwen-VL OCR, Qdrant retrieval, DINOv2 and ArcFace, deterministic fusion, guardrails and explicit unknowns.
Field-Sales & Commercial Operations Platform
Offline-first platform for field sales reps with bidirectional 1C ERP integration, KPI, management BI, forecasting and merchandising AI. 98% of the company's ERP orders went through it.
Supply Planning Action Agent
Embedded self-hosted Qwen agent with 11 bounded tools: planning actions and UI navigation, domain resolution before any change, and user identity injected by the server.
More projects
AI Infrastructure Control Plane & MCP Gateway
Monitoring for GPU, models and services with AI-specific detectors (LLM endpoint drift, silent OCR degradation), an incident assistant and a least-privilege MCP gateway.
Jarvis — Voice Multi-Agent Orchestrator
Real-time voice interface over a fleet of agents: WebRTC, agent registry and command queue, role-based capabilities and allowlisted browser actions.
Offline Merchandising Terminal
Native Android field app (Kotlin, Compose, Room, WorkManager) that captures shelf photos offline and feeds the recognition pipeline without blocking on GPU inference.
Internal Developer Platform & Release Orchestration
Delivery control plane for multi-environment releases: repository state, selective cherry-pick promotion, conflict handling, dependency preflight, build and health verification.
AI Marketing & Brand Growth Platform
Multi-source marketing intelligence for a consumer beverage brand: website, Instagram and search analytics, content performance and influencer vetting.
Fitness Coaching Platform (prototype)
Full-stack coach/client prototype on Next.js and Payload CMS with private signed media, RBAC, chat and 204 automated tests including security suites.
Skills
Contact
Open to applied AI, computer vision, RAG and LLM-agent roles — in Tbilisi or remote.