Shah Khurram

Lead AI engineering across 5 product verticals; set org-wide standards for multi-tenant isolation, plugin extensibility, and security hardening (encryption, SSRF guards, rate limiting) while mentoring 10 engineers. Built a visual agent-orchestration platform React Flow node-graph builder on a streaming FastAPI engine letting teams compose agents from typed nodes (LLM, router, RAG, MCP tool, human-in-the-loop) so non-AI engineers ship AI features. Implemented a native MCP client (Streamable-HTTP transport, JSON-RPC handshake, SSE) enabling agents to call tools on any MCP server; connection layer spans 4 databases, 5+ vector stores, and 7 LLM providers. Designed a credential-isolation architecture nodes reference connections by ID, secrets never leave an external store enabling safe multi-tenant tool access in embeddable deployments. Established platform-wide AI cost governance: an OpenTelemetry GenAI token/cost ledger with cache-aware accounting and per-tenant attribution, delivering per-model unit economics on 3.5M tokens/month. Contained single-vendor outages behind LLM reliability infrastructure (provider-scoped circuit breakers, error classification, fallback dispatch), alongside shared RAG primitives and HIPAA-regulated document-AI pipelines with regression-tested output.

Engineered multi-agent systems with autonomous task execution and role-based coordination spanning 12+ enterprise workflows across 4 internal operations teams. Architected dual-layer vector memory agent-specific Pinecone stores plus a global cross-agent knowledge base with local-first retrieval and global fallback. Shipped real-time voice AI assistants with low-latency streaming STT/TTS (~500 ms end-to-end) and per-agent voice, personality, and model configuration. Owned AWS infrastructure end to end (EC2, RDS, S3 security, monitoring, autoscaling, cost optimization) and GitHub Actions CI/CD with automated rollbacks across staging and production. Integrated Stripe (subscription + usage-based billing), Tavily search, Cohere embeddings, Postmark, and Google OAuth 2.0; mentored developers and represented engineering in weekly CEO-level strategy meetings.

Built an Intercom-style agentic AI platform multi-agent support, per-agent vector knowledge bases, and embeddable JavaScript chat/voice widgets deployed into client SaaS products. Replaced a manual data-entry team with privacy-first document intelligence (PDF, DOCX, OCR) that auto-fills Hospital Management Systems using on-premises Hugging Face LLMs 5,000 documents/month with zero PHI leaving the network. Developed backend systems in PHP (Laravel, CodeIgniter) and Node.js powering appointment scheduling and workflow automation for healthcare clients, with real-time WebSocket communication. Implemented recursive web-scraping pipelines with depth-controlled crawling and automated vector-database ingestion, powering semantic search and RAG over continuously refreshed client corpora. Promoted from Trainee Developer to full-time engineer within 3 months while completing B.Tech under university no- objection certification.