Quiet Clairvoyance

Foresight you earn in hindsight.

Building AI Agents That Matter

I recently completed the “Building AI Agents for the Enterprise” course led by Abhijith Neerkaje (Head of Data Science) and Ajay Shenoy (PhD, IISc) — one of the most rigorous, practical, and engineering-first AI courses out there.

For my capstone, I built TBawareGPT: a multilingual Q&A + triage agent for frontline Tuberculosis community healthworkers. It’s designed to be medically safe, constraint-driven, and ready for real-world use.

Five things that stood out:

  1. Engineering-first: Predictable agent behavior, explicit safety boundaries, modular architecture, and production-ready patterns.
  2. Applied impact: Reliable & optimized agents for field, symptom guidance, counseling support, and bilingual explanations.
  3. Multi-channel interfaces: CLI for dev/eval, a scalable REST API, mobile-first (Telegram), and a lightweight web UI.
  4. Domain safety by design: Strictly aligned with WHO TB & RNTCP/NTEP guidance, “suggest, not prescribe” policy, and misuse guardrails.
  5. Evaluated, not just shipped: Robust test harnesses, memory & summarization, Postgres history, and full audit trails.

Full-stack architecture highlights:

  1. Agent Core: LLaMA-3.3-70B (Groq), medical guardrails, caching, and real-time English ↔ Hindi localization.
  2. Medical Knowledge & Tools: WHO/RNTCP-compliant Q&A, TB-type inference (DS/MDR/XDR/Pediatric), “suggest not prescribe” regimen guidance, and treatment locator.
  3. Memory & Context: Conversation buffers, PostgreSQL memory, user-scoped agents, CSV session logs.
  4. Infrastructure & Observability: Fully containerized agent/deploy/evals, model pre-caching, Docker profiles, plus tracing/APM (roadmap).
  5. Evaluation & Learning: Domain harness with 40+ TB scenarios, semantic scoring, latency/stability profiling, and continuous prompt/safety refinement.

My key takeaway: the true frontier of AI will emerge where reasoning meets irreversible consequence — medicine, robotics, defence.

If you’re building AI infrastructure or safety-critical workflows, the course teaches the skills, mental model, and design patterns to build systems you can trust. Completing this course was a key part of meeting my yearly professional development goal.

Highly recommended for anyone serious about operating AI agents in the field.