Mohammad Kashif

About

Built and deployed a compliance-sensitive LLM pipeline that extracts rebate logic from hospital–supplier medical contracts, using multi-layer redaction (regex/NER/dictionary), schema-validated structured outputs, and secure placeholder reinsertion. Hosted models on Amazon Bedrock to avoid retraining on sensitive data and improved both accuracy and cost by splitting the workflow into a lightweight section classifier plus a fine-tuned extraction model, orchestrated with LangChain and evaluated via layered, test-driven agent assessments.

Experience

Education

Key Strengths

• Built and deployed an LLM system for extracting rebate logic from sensitive medical contracts
• Strong compliance-first architecture (multi-layer redaction + secure placeholder reinsertion + schema validation)
• Pragmatic model strategy: decomposed a single large-model pass into classifier + extraction stages to improve accuracy and reduce cost
• Hands-on orchestration experience (LangChain/LangGraph/custom multi-agent)
• Systematic reliability approach using layered evaluation, labeled data, and failure-driven test expansion
• Applied parameter-efficient fine-tuning with high-quality supervised samples to stabilize extraction performance
• Effective collaboration with non-technical stakeholders via user feedback loops to improve agent communication
• Designed and built low-latency AI voice calling system integrating Twilio + OpenAI with post-call dashboard reporting
• Pragmatic tradeoff-making to reduce complexity while maintaining UX, latency, and cost constraints
• Led backend refactor with backward-compatible schema, validation, and reconciliation to preserve data integrity
• Risk-managed incremental rollout using parallel logic runs and feature flags
• Implemented auth and row-level security in production using Clerk + Supabase RLS with automated testing
• Proactively identified missed edge cases in document ingestion (partial eligibility and timing boundaries) and added pipeline checks
• Built an AI calling agent integrating OpenAI Realtime API with Twilio Streaming
• Strong prompt/agent behavior tuning via role definition and constraints
• Pragmatic MVP scoping and prioritization under time/resource constraints
• Architected task-oriented agent workflow optimized for B2C versatility
• Led cross-functional alignment on shared architecture and backend schema to fix fragmented data flow
• Influences without authority through documentation and stakeholder walkthroughs