About
About
I’m Jyoti Ranjan — a GenAI architect with 19+ years spanning enterprise development, solution architecture, and production AI systems. I build things that have to work in the real world: under latency budgets, under compliance regimes, and under the scrutiny of people whose job is to say no.
What I’m focused on now
My current work sits at the intersection of agentic AI and security. As enterprises move AI agents into production, two questions get sharp fast: can this agent be trusted to act on our behalf, and can we prove to an auditor that we checked? Most of the industry is answering the first with scanners. I’m interested in the harder second half — turning a security finding into compliance-grade evidence — which is what I’m building toward with Provenire, an open-core MCP security and compliance-evidence tool.
Alongside that, I design and ship production LLM systems: a low-latency voice AI agent, retrieval and multi-agent architectures, and Document AI pipelines — including work on low-resource languages that most models ignore.
Background
- US patent holder and arXiv co-author, with roots in low-resource-language AI (Odia OCR) and an IIT research track.
- A picks-and-shovels bias: I’d rather build the infrastructure other people depend on than the thing on top.
Core expertise
Agentic & LLM systems — multi-agent orchestration (LangGraph, CrewAI, Autogen), RAG / GraphRAG, LLM fine-tuning (SFT, RLHF, DPO), vLLM inference, guardrails (NeMo).
AI security & compliance — MCP/agent attack-surface analysis, framework-neutral finding→control mapping, evidence generation.
Voice & Document AI — real-time voice agents (LiveKit · Deepgram · ElevenLabs), OCR and layout-aware extraction (Donut, TrOCR, Qwen2.5-VL).
Platform — Python · Azure OpenAI · AWS Bedrock/SageMaker · Docker · Kubernetes · Neo4j & vector DBs · MLflow.