Technology manager designing agentic AI that regulated teams can audit. My agents check their own work with deterministic gates, run evals as CI quality gates and stop for an independent human sign-off, with audit trails end to end. 17+ years delivering for life sciences and BFSI clients. Claude Certified Architect – Foundations. Chennai, India.
Open to AI architect and AI solutions roles (remote or Chennai) · available to join immediately.
LangGraph · MCP · RAG · Multi-model LLMs · Amazon Bedrock · Azure OpenAI · LLMOps & evals · AI governance · Python · R · SAS
Portfolio, reference architecture and all demos: grprakasam.github.io · 4-minute app walkthrough
1-minute intro: my resume as a short animated walkthrough (captions, no audio).
| Project | What it does | Evidence |
|---|---|---|
| ADAM AI | LangGraph agent that turns a CDISC ADaM spec into an R program, runs it in a network-isolated sandbox, checks it against an independent reference and stops for independent human sign-off. Hash-chained audit log, cheap-first model routing, MCP server. | Pharmaverse ADSL benchmark in CI: 28/28 variables, 306 subjects, exact match after one repair loop (replay) · demo · app walkthrough |
| agentic-ai-learning-lab | Two tracks. Built: converts legacy XML/JSON workflows into Power Automate drafts with deterministic rules first, a bounded LLM with guardrails second, everything else flagged. Learning: RAG with retrieval evaluation on AWS with Terraform (design done; build planned from Nov 2026). | Scorecard: 82% of actions converted by rules alone, zero silent failures in 29 behavioral tests (offline, synthetic) · demo |
| AI Exam Builder | The LLM writes questions and feedback; code decides scores, difficulty and SM-2 mastery. Access-controlled attempts. | Integration tests and build in CI · demo |
- Model output is untrusted until a deterministic check passes.
- Cheap model first, stronger model on repair, a person last.
- Every capability states its evidence: unit-tested, replayed, mocked or live.
- Significant decisions are written down as ADRs (ADAM AI's ten).
Portfolio page · LinkedIn · Claude Certified Architect – Foundations (2026) · PharmaSUG US 2022 co-author (R Shiny in clinical programming)



