Senior AI/ML Engineer & AI Architect

Multi-agent corporate finance audit & fraud investigation system — investigates suspicious payments, duplicate invoices, ghost vendors, and policy breaches, with human approval gates, audit logs, and report export. Public hackathon project with a live demo.

Multi-model AI chat where OpenAI, Claude, and Gemini agents answer the same prompt, then compare, critique, and debate each other to converge on a stronger, more reliable answer. Routes one question across providers in parallel, surfaces where models disagree, and uses Playwright for automated end-to-end testing of the chat flows.

Safe AI analytics assistant for PostgreSQL using approved SQL templates, RAG, embeddings, and admin-focused query workflows.

Research and prototype work for Uzbek voice AI using STT, TTS, realtime latency optimization, and multilingual AI assistant design.

Plant disease detection and agricultural AI pipeline using computer vision, MLflow, model validation, and dataset workflows.

AI-powered accessibility SaaS that generates WCAG-compliant alt text from images and runs site accessibility audits. Built on Laravel 12 + Livewire with Gemini 2.5 Flash multimodal vision, a credit-based usage system, Paddle subscription billing, OAuth login, and a Sanctum-protected REST API for third-party integrations (e.g. WordPress).

Searchable web directory of AI agent skills discovered from GitHub and other public sources. End-to-end data pipeline — discovery → extraction → PostgreSQL → Gemini enrichment → API → UI — with a Next.js frontend and FastAPI backend, fully Dockerized.

FastAPI prototype that compares AI agronomy-assistant architectures for plant & crop diagnosis. Runs three pipelines side by side — pure vision LLM, hybrid RAG (Postgres FTS + pgvector), and RAG + reranker — with an image-quality preflight, RU / Uzbek / mixed-language normalization, confidence fusion, safety validation against unsafe dosage advice, and an evaluation runner. Gemini vision; Dockerized.

Autonomous multi-agent system that coordinates the full WordPress-plugin lifecycle — market-intelligence and opportunity scoring, development, marketing, distribution, and support — through specialized Claude/GPT agents with a monitoring dashboard. Market-research and opportunity-scoring agents are active; build/QA agents are in progress.

Runnable plant-diagnosis pipeline that takes a leaf photo plus a voice message, transcribes the audio and auto-detects Russian/Uzbek, then sends image + transcript as a single multimodal Gemini request. Returns structured JSON — diagnosis, confidence, recommendations, crop match, and human-review flags — with graceful fallback JSON on model failure and a test suite covering invalid images, low confidence, crop mismatch, and chemical-advice safety gating.