GeorgiKirilov
Georgi
Work
Recent large-scale platforms I've built and scaled in production.
Serving a platform of 5 million creators, 20,000+ of them daily. Built and scaled the infrastructure handling GPU fleet management, real-time generation queues, and high-traffic serving — maintaining sub-second response times under heavy concurrent load. Took each new model from evaluation to production, benchmarking output quality against cost per inference.
Main engineer — designed, built and run the full platform. Dual-vector semantic search in Qdrant across 2 million images, with an MCP interface so agents can query the index directly. Users find visually similar content through natural language and image-to-image queries.
freegen
Millions of AI-generated images served. Built a high-throughput inference platform with autoscaling and self-healing mechanisms using GPU compute on local infrastructure, Modal and Vast.ai — optimized for scale and cost-efficiency with distributed pipelines that handle burst traffic without over-provisioning.
Real Estate Tracker
Paste a listing URL into a Telegram bot and get an instant read on whether the flat is over- or under-priced for its neighbourhood — percentile rank and a % gap against a rolling 90-day median, plus 3/6/12-month trend and active inventory. Comparisons are made like-for-like: prices are VAT-normalised, a vision model grades finish from the listing photos, and shell or partially-finished flats carry an estimated fit-out cost before being benchmarked. Perceptual hashing recognises relisted properties instead of counting them as new supply. In production since July 2026.
Dual-process Python system — an autonomous daemon executes strategies on 4h candle boundaries while a custom MCP server exposes the full control plane to Claude Desktop for natural-language oversight. Pluggable strategy engine backed by Coinbase Advanced Trade API, with shared state across processes and live Telegram notifications. Currently operational.
Expertise
01
LLMs that do things rather than just answer — tool-calling pipelines, MCP servers exposing real control planes, and autonomous daemons with human oversight. Built for systems where a wrong action costs money.
02
Integrating large language models into production products — prompt engineering, fine-tuning pipelines, RAG architectures, and reliable inference at scale. From prototype to shipped feature, fast.
03
End-to-end design and deployment of AI compute infrastructure — GPU fleet orchestration, auto-scaling, and cost optimization across cloud and bare-metal providers. Built for high availability and production-grade reliability.
04
High-throughput inference pipelines for image and video generation — queue management, batching strategies, and latency optimization for demanding workloads. Millions of generations served across multiple platforms.
In the fast-moving world of AI, I ship infrastructure that holds under pressure — GPU fleets that don't melt, pipelines that don't drop jobs, and integrations that actually make it to production.
Open to consulting, collaborations, and interesting projects in AI infrastructure and Systems Design.