I turn business and product problems into AI systems — from model selection and inference strategy to agents, memory, retrieval and deployment.
Business teams often know exactly what needs to improve but not which AI approach makes technical or economic sense. AI engineers understand the underlying models and infrastructure but may not own the product, user, or business context. I connect the two.
Concrete systems built around hard memory, privacy, hardware, and retrieval constraints.
A local-first memory engine exploring how retrieval, provenance and long-term memory can make AI systems more capable.
The synthetic graph win (0.325 → 0.800) failed to transfer to real dialogue: 0.168 recall@10 on LongMemEval vs 0.193 BM25 baseline.
An on-premise realtime communication system designed around hard privacy, hardware and infrastructure constraints.
Rockchip RK3528A/RK3588S appliance with SoC-ID + eMMC-CID binding, 33-byte binary wire protocol, and a zero-knowledge AI sidecar over unix socket.
┌───────────────────────────────────────────────┐
│ CLIENT LAN (ON-PREM) │
│ ┌─────────────────────────────────────────┐ │
│ │ ARM APPLIANCE (Rockchip RK3528A / 2GB) │ │
│ │ ┌────────────┐ ┌────────────────┐ │ │
│ │ │ Axum WS │────▶│ SQLCipher eMMC │ │ │
│ │ │ Blind Hub │ │ Write-Buffered │ │ │
│ │ └──────┬─────┘ └────────────────┘ │ │
│ │ │ (Unix Socket) │ │
│ │ ▼ │ │
│ │ ┌────────────┐ (Zero-Knowledge) │ │
│ │ │ AI Sidecar │ │ │
│ │ └────────────┘ │ │
│ └─────────┬───────────────────────────────┘ │
│ │ WSS (Binary Protocol v2) │
│ LAN PWA / Swift iOS Clients (WASM/UniFFI) │
└───────────────────────────────────────────────┘The layers I work across — from model selection to product adoption.
Running models, not just calling APIs. Hands-on experiments with open models, inference runtimes, hardware limits, and deployment trade-offs.
Published research notes, benchmark reconstructions, and empirical investigations.
A typed relation graph raised synthetic multi-hop recall@10 from 0.325 to 0.800. On a small LongMemEval sample, it fell below BM25. An early research note on benchmark-shaped wins, real dialogue, and what failed next.
Before working deeply with AI systems, I spent nearly a decade in design, product and creative leadership.
That background became the other half of my current work: understanding users, business constraints, communication, information systems and product experience. It is why I don't build tech for tech's sake.
The recent AI wave pulled me from creative leadership into models, inference, agents and AI systems R&D.
Today my strongest skill sits between two worlds: understanding what a business needs, understanding what current AI technology can actually do, and designing the system that connects them.
I work heavily with coding agents for implementation and rapid iteration. My ownership is in problem framing, architecture, experiments, evaluation and product decisions, while I continue developing deeper independent engineering fundamentals.
International remote teams, roles with relocation potential, and selected high-impact opportunities in Tashkent.