Independent AI/ML engineer · Cost-efficient & on-prem LLM infrastructure · AI governance · Tashkent, Uzbekistan

I build open-source infrastructure for running LLMs affordably and safely — cost routing & metering, on-prem/local-first inference, model evaluation, and agent governance. AMD Developer Hackathon ACT I winner (REPOMIND); REPOMIND v3 submitted to ACT II. Latest OSS: Frugal — the layer that keeps agent inference cheap, local, and verified.

REPOMIND v3 — a repo-scale AI coding agent that runs entirely on your own hardware, plus an LLM cost-router that keeps token spend down. Built for regulated / air-gapped teams the cloud isn't allowed to touch. Submitted to the AMD Developer Hackathon ACT II (Track 3 · Unicorn).

• 🧠 256K-context repository understanding on a single AMD MI300X (FP8 KV cache)
• ⚡ Measured: 31/31 parallel users @ 8K–64K · $4.12 total compute (124 min) · −89–90% cost via the router
• 🥇 Builds on REPOMIND, which WON 1st place at AMD Developer Hackathon ACT I (AI Agents + Outstanding Social Engagement) — prize: AMD Radeon AI PRO R9700 GPU + cash
• 📋 ACT II status: qualified in the automated pre-screen — the human panel decides the final result (not yet won)
• ★ Repo → SRKRZ23/repomind-v3

Frugal — run AI agents cheap, local, and verified. Meters every model call, routes the cheap/local model first, and escalates to a strong model only when a real check fails — then proves quality in CI. Inference is now ~85% of the enterprise AI budget (Uber burned its whole 2026 coding budget by April; one firm spent $500M on Claude in a single month) — Frugal routes around it.

• 📊 Measured on a real cluster: a 3B model matched a 14B on 83% of hard tasks, ~4.7–11× faster → ~75–97% cheaper
• 🧪 56 tests, CI green, 0 runtime deps · real bugs found & fixed by our own pressure / fuzz / concurrency suites
• 💸 Modeled enterprise saving: at $10M/mo inference spend → $60–90M/yr (roster + math on the live site)
• 🌐 Live site + interactive deck · ★ Repo · Apache-2.0

                ┌───────────────────────────────────┐
                │  CITADEL — Model Evaluation       │  ← evaluates which models are safe
                │  github.com/SRKRZ23/citadel        │
                └─────────────┬─────────────────────┘
                              │
        ┌─────────────────────┼─────────────────────┐
        │                     │                     │
┌───────▼────────┐  ┌─────────▼─────────┐  ┌────────▼────────┐
│  FORGE         │  │  SOUF AI          │  │  ATLAS          │
│  Policy Gen    │  │  Inline Firewall  │  │  Agent Routing  │
│  (IBM Bob)     │  │  (Veea Lobster T) │  │  (multi-agent)  │
└────────────────┘  └───────────────────┘  └─────────────────┘
                              │
                  ┌───────────▼───────────┐
                  │  Shared Ed25519       │
                  │  audit chain          │
                  │  (built once, reused) │
                  └───────────────────────┘

60%+ cross-product code reuse. One codebase. One mission: governance infrastructure that's free, open, and reproducible.

I shipped 4 hackathon submissions in 3 days (May 17–19) via shared infrastructure earned from prior submissions:

• 🔥 AMD MI300X GPU access (earned at REPOMIND, AMD Hackathon May 11) → real Gemma 3 27B benchmark in CITADEL (87.5% authority resistance, 72.8 tok/s on 192GB HBM3)
• 🔧 IBM Bob coding agent credits (40 Bobcoins from IBM Bob Hackathon) → 27 to FORGE + 13 to CITADEL extension (multilingual prompts, edge adapters, compliance docs)
• ⚙️ SOUF AI Ed25519 audit chain → reused as CITADEL L7 provenance layer + ATLAS ingress audit core

Each prize funds the next submission. Each architecture decision pays for two more products.

F1=1.000 on 231 adversarial prompts across 5 benchmarks. 1,800× faster than Meta Prompt Guard 2 (measured 0.079ms vs 92.4ms on A100 per Meta's model card). Built-in HIPAA + PCI-DSS packs. Ed25519-signed audit chain. Video walkthrough · Lablab submission

13-layer L0–L12 architecture. Gemma 4 27B benchmarked against 5 frontier models. EU AI Act / NIST / ISO 42001 / HIPAA compliance reports. Real AMD MI300X pilot run. Live Streamlit demo · Video

OWASP LLM Top 10 mapping → YAML policies → BobShell audit trail. Built with IBM Bob coding agent. Video walkthrough

6-layer governed pipeline: Voice → SOUF AI DPI gate → Gemini orchestrator → Featherless router → Tool executor → Ed25519 audit. 29/29 tests PASS in under 1 second. All 5 Milan AI Week sponsors integrated (Speechmatics + Featherless + Gemini + Vultr + Kraken). Video walkthrough · Lablab submission

Open-source large-context (256K) FP8 inference on AMD MI300X. WON 1st place — AMD Developer Hackathon ACT I (AI Agents track) + Outstanding Social Engagement. ACT II (v3: on-prem agent + LLM cost-router) submitted, qualified in the automated pre-screen.

Quantitative framework for measuring when LLMs surrender independent reasoning under authority pressure. 7 frontier models, 2,520 measurements. Wilson 95% CI methodology. Zenodo DOI: 10.5281/zenodo.19791329

• Founded UCAR (2023, Tashkent) — U-Start Demo Day 1st place + 85M UZS grant; Korea-Uzbekistan Startup Exchange 2nd place in Seoul
• Youngest student in UJC's PMP-43 senior executive cohort
• Scholarship to Zhejiang University
• Foundation Year at Leeds Beckett University, UK
• Kaggle SPR Mammography: 7/371 (Top 1.9%, Medical AI)
• CVPR 2026 CV4CHL Gait Challenge — Kaggle LB 2nd (0.90271); official 6th on the held-out set; Proceedings co-author
• AMD Developer Hackathon 2026 Track 1 with REPOMIND

• ORCID: 0009-0007-0731-4247
• Zenodo: 10.5281/zenodo.19791329
• HuggingFace: @ZeroR3
• LinkedIn: sardor-razikov-569a5327b
• X / Twitter: @SardorRazi99093
• Email: razikovsardor1@gmail.com

"Every researcher deserves the same evaluation infrastructure OpenAI has internally. Free. Open. Reproducible."

— Built solo from Tashkent · MIT licensed · Available for strategic conversations.