Abubakar Omer

Consulting ML Engineer delivering production AI systems for Optum Serve, a federal healthcare client.

• Multi-Agent Analytics Platform: Architected a multi-agent platform with an LLM-based intent classifier coordinating two specialized agents: a RAG retrieval agent over pre-generated reports and a DuckDB code-execution agent, enabling business users to query operational data conversationally without SQL or code.
• SAS-to-Python Modernization Agent: Engineered an agentic SAS-to-Python pipeline using Claude Code CLI with a custom MCP server for documentation upload, GitHub push, and Databricks job execution, generating complete Python packages with tests and docs in a single invocation.
• SQL-RAG Schema Classifier: Built a BERT-based multilabel classifier identifying schema columns from natural language queries. Synthesized 200k+ labeled examples via Azure OpenAI, tracked via MLflow on SageMaker, achieving 86% average per-label accuracy.
• Automated Thematic Analysis: Designed a dual-stage metadata-driven LLM pipeline using fine-tuned GPT-4o on Azure OpenAI with multi-level parallelism, reducing survey processing time by 75% (30+ mins to <7 mins).
• Form Type Classifier: Developed a CPU-only, training-free classifier for scanned federal medical forms using one-shot ORB feature matching and RANSAC alignment, achieving 95–98% confidence across 21 form types with a single exemplar per class.
• VA Claims ACE Pipeline: Delivered a 4-phase multi-agent pipeline on AWS Bedrock: a preprocessing agent indexes PDFs into a tagged Knowledge Base; five domain-specialized agents perform tag-filtered retrieval with PydanticAI outputs; a Conflict Detection Agent flags inconsistencies; a rules-based ACE Determination Agent routes to human review, approval, or C&P exam.

Consulting role providing data science solutions across pharmaceutical, government, and enterprise clients

Client: Merck Pharmaceuticals

• Cell Viability Prediction: Developed real-time cell viability monitoring system using custom YOLO object detection models, achieving 92% accuracy in collaboration with laboratory scientists to automate previously manual video analysis processes.

Client: US Census Bureau

• Multi-Label Text Classification: Trained a custom BERT-based multilabel token classification model for race coding across a dataset of 2M+ entries, automating 95% of manual data coding tasks.

Internal Projects

• GenNet AI: Launched an AI-powered physician platform for RAG-based querying over consultation transcripts and patient data with native OpenEMR integration; advised physicians and clinical stakeholders on GenAI adoption roadmap and self-hosted LLM deployment strategy, deploying fully on vLLM and h2oGPT.
• Churn Prediction Platform: Brought a modular churn prediction pipeline from prototype to production on AWS Step Functions with ECS jobs, serving multiple clients simultaneously. Used PyCaret for automated model selection and SHAP for explainability, achieving 25% reduction in churn rates.