Ilya Khalafi

• Achieved a score of 80% on an in-house agent built for a Japanese partner company, retrieving facts from 1.5M+ chunks using knowledge graphs, Qdrant hybrid search, Elasticsearch keyword search, and other tools.
• Scored 96% for another client by converting chart- and table-heavy documents into SQL tables and Qdrant hybrid indexes, answering time-interval questions with a Qwen3.6-35B model on DGX Spark hardware.
• Currently fine-tuning Qwen3.6-35B with SFT + LiPO + GSPO using the verl library on a cluster of 8× H200 GPUs for that agent.
• Implemented a synthetic data generation pipeline spanning simple conversational, internal security, hallucination-check, SQL, and search-grounded categories with k rollouts for SFT/LiPO; SFT phase stable with MiCA initialization, testing LiPO before DAPO/GSPO.

Worked on OneMan-AI, a business intelligence AI agent for analyzing heterogeneous, islanded databases and files across enterprise environments.

• Implemented MCP tools on top of Trino to integrate multiple data sources — from SQL and NoSQL databases to Excel files — through a unified query layer for the AI agent.
• Built highly efficient workflows for ingesting structured and unstructured files, and deployed the data engineering pipeline with Hatchet for production file-processing workflows.
• Developed the agent's schema exploration and intent recognition capabilities, plus key agentic workflow and backend components.
• Deployed and maintained platform services including GitLab, GitLab Runner, Langfuse, and Portainer, with CI/CD pipelines for all microservices.
• Implemented most of the backend services with Django and the web UI with Next.js for end-to-end delivery of the analytics platform.

• Improved pipeline throughput 2× by designing batching and asynchronous I/O operations for large-scale document processing.
• Achieved near-linear CPU scaling with a fault-tolerant multi-process architecture handling a 200K+ document corpus efficiently.
• Constructed asynchronous pipelines for ontology and knowledge graph extraction with page-level reference integrity.
• Engineered the asynchronous AI agent core on LangChain / LangGraph with efficient memory management and evaluation hooks.
• Conducted comprehensive benchmarking of vector database search relevance and latency for enterprise use cases.
• Enhanced agent observability with Langfuse-based multi-metric tracing and automated performance dashboards.