Prakasam R

I design agentic AI that regulated teams can audit.

Technology manager with 17+ years across life sciences and banking. My agents check their own work with deterministic gates, run evaluations as CI quality gates, and stop for an independent human sign-off, with an audit trail end to end.

Claude Certified Architect – FoundationsChennai, India · open to remoteAvailable to join immediately

Watch the 4-minute app walkthrough LinkedIn profile GitHub

My resume in one minute: an animated walkthrough with captions, no audio.

How I architect an agent

ADAM AI is my reference design. Each stage is a decision I wrote down as an architecture decision record (ADR), and each one exists because a regulated reviewer will ask about it.

  1. 1Spec inADaM specification, YAML or Excel
  2. 2GenerateLLM drafts the R program; cheap model firstADR 0004
  3. 3Static gatessafety, syntax, variable coverage, CDISC namingADR 0002
  4. 4Sandboxed runR in a container with no networkADR 0009
  5. 5Compareevery variable against an independent referenceADR 0006
  6. 6Human sign-offan independent reviewer; the requester cannot approveADR 0003

Repair loop. Any failure at stages 3 to 5 goes back to stage 2 with the evidence: the error, or the mismatched variables with example subjects. Repeated failures escalate to a stronger model, then to a person.

What I have built, and how each number was measured

All three are proofs of concept on public or synthetic data. Each demo replays a real run, and each stamp says how that number was produced.

ADAM AI

Proof of concept on clinical data

A LangGraph agent turns a CDISC ADaM specification into an R program, runs it in a network-isolated sandbox, compares the output with an independent reference and stops for a person to sign off. Every step goes into a hash-chained audit log.

  • 28 of 28 ADSL variables matched306 subjects, public pharmaverse benchmark, rerun in CI on every change
    Replay with a known-good program
  • About 2 days to about 20 minutesfirst-draft programming for ADSL and ADAE in copilot mode
    Internal trial, self-measured

Workflow modernization agent

Track 1 of my agentic AI learning lab, built and tested offline

Converts legacy SharePoint- and Nintex-style workflows into Power Automate drafts. Deterministic rules go first and a bounded LLM second; anything left over is flagged for a person. A differential test then checks that each draft behaves like the original.

  • 0 silent failures in 29 behavioral testsevery difference traces back to a step that was already flagged
    Offline run, 12 synthetic workflows
  • 82% of actions converted by rules alone95% with the bounded LLM step, using hand-written stand-in replies
    Fixtures, not a live model

Learning next: regulatory RAG over public FDA guidance on AWS with Terraform, with retrieval evaluation and an answer-and-verify flow. Design done in October 2026; build planned from November 2026.

AI Exam Builder

Proof of concept for exam practice

An LLM writes practice questions and feedback for eight exam standards, including NEET and JEE. Code decides scores, difficulty, spaced-repetition reviews and access, so a wrong or unavailable model cannot change a student's result.

  • Bad model output rejected before savingscores computed by code and kept during a feedback-model outage; repeat submissions refused
    API run, stand-in model replies
  • 8 integration tests and a frontend buildrun in GitHub Actions on every push
    CI

Background and credentials

Before AI, I delivered clinical analytics and data work for Pfizer, Johnson & Johnson, Abbott, Baxter and Vestas, and analyst, BI and QA work for State Street, American Express, GE Corporate Treasury, NewDay and Standard Bank of South Africa. At Ephicacy I led a team of 10 (3 direct reports) and trained 200+ colleagues in R and Shiny.