Ananthaprakash Thiyagarajan

Enterprise AI has a trust problem. Models hallucinate, pipelines fail silently, and outputs can reach people with no clear audit trail. At Deloitte, my work focuses on that gap for clients in regulated environments.

I design multi-agent orchestration systems using LangGraph, LangChain, and agent development frameworks. The goal is not only better answers. The goal is traceable reasoning, confidence signals, escalation paths, and graceful failure when a system reaches the edge of what it knows.

The work spans LLMOps infrastructure, responsible AI guardrails, client AI strategy, PII handling, evaluation design, and the translation layer between model capability and stakeholder trust.

Makerble is a CRM and impact-measurement product used by charities and social organisations to track outcomes and report results. The challenge was a credibility gap: organisations were doing meaningful work across the UK, Europe, and Africa, but turning that activity into proof was manual, slow, and hard for funders to trust.

I built the intelligence layer that made impact easier to explain. RAG-powered agentic pipelines ingested contribution data and produced grounded, tone-aligned insights at dashboard speed. Conversational AI built with Google CCAI helped customers understand the platform and helped the sales team move complex deals forward.

Underneath that experience, I worked on distributed Golang microservices after the Rails monolith migration and scaled the product’s infrastructure to 50M+ monthly requests. I also contributed to OCR document pipelines that fed Makerble’s data layer and improved the quality of downstream insight generation.

There was no team, no platform, and no process. There was a mission: bring quality dialysis care to patients in tier 2 and tier 3 cities who could not reliably access it.

I built the engineering organization from four people to seventy-five in three years. I also led the clinical AI platform: multimodal models trained on vitals, lab reports, and clinical notes, using LSTMs and Transformers to identify complications before they became crises.

The inference layer ran across edge devices and cloud AI platforms with sub-200ms prediction latency. In a clinical setting, latency is not an engineering vanity metric. It changes how quickly a care team can respond.

The proprietary risk assessment model is still running. Nearly 10,000 patients benefit from it every day. That number is the one I return to when deciding what work is worth doing.