// 01 · profilesaketh ram billa

Hi, I'm Saketh.

I build AI for early-stage startups. Mostly agents and RAG. Mostly the unsexy stuff that makes them work in prod.

Saketh Ram
basedDelhi, India · UTC+5:30
currentFounding AI Engineer @ Turgon.AI
shipped5+ 0→1 products, still counting
stackLangGraph · Claude · FastAPI · Next.js
educationBTech CSE · IIIT Pune ('24)
fuelcoffee · coke · 2am ships
// 02 · experience2 roles · 4 shipped

Where I've shipped.

The titles, but also what actually got built.

Oct 2024 — present

Founding Engineer

Turgon.AI · New Delhi · backed by Founder Collective

▸ TAWP — Browser Automation Agent

  • Built the whole thing from zero — a LangGraph Plan-Act-Observe loop with Claude and GPT-4 driving Playwright. Point it at a portal, give it credentials and a plain-English goal, it figures out the rest and runs unattended.
  • Added a recipe cache that remembers successful execution paths per portal — repeats run 5-6× faster. Also solved MFA/TOTP, the S3 file handoff, and the audit trail you need when you're running someone else's credentials in prod.
PythonFastAPILangGraphClaudeGPT-4PlaywrightS3

▸ Clubs CMS — Multi-tenant AI Platform

  • Architected a multi-tenant platform for 3+ branches — tenant isolation, Supabase JWT RBAC, PayU digital wallet with automated reconciliation + analytics dashboards. Not just the happy path; the full operational surface.
  • Wrote the comms engine — cron-driven push notifications, trigger-driven email flows (RSVPs, renewals) via Twilio SendGrid. All configurable from the CMS so non-engineers can edit the logic without calling me.
  • Built the in-CMS analytical chatbot — multi-agent LangGraph flow with a FastAPI backend, HiCharts visualizations embedded directly. Branch managers ask questions in English, get charts back.
  • Tuned performance end-to-end — Redis caching, Prisma indexes where they mattered, RTK Query keeping CMS + mobile + API in real-time sync.
Next.jsPrismaSupabaseRedisRTK QueryPayUSendGridLangGraphFastAPI

▸ The Quorum — White-label Mobile App

  • Built the mobile platform on a BFF architecture — multi-role auth (club member / check-in scanner), CMS-driven content that updates in real time on devices.
  • Shipped the QR-based club access and RSVP flow end-to-end — event bookings, push/pop notifications, PayU payments. The app door staff uses when you walk into the club.
  • Embedded an AI chatbot with LangGraph + Whisper. Members can voice- or text-ask about events, reciprocal clubs, perks, venues, private hires — the questions that don't need to reach staff anymore.
React NativeExpoLangGraphWhisperPayUTwilio 2FA
May — Jul 2023

ML Research Intern

TCS Research & Innovation · Pune

  • Trained an emotion-detection model on low-quality phone-camera images — the kind papers never benchmark on because they're messy. Hit 95% accuracy after benchmarking a dozen CNN architectures + training regimes against actual noise.
  • Shipped the end-to-end pipeline — React frontend, server-side API integration. Not just the notebook.
PythonPyTorchOpenCVCNNsReact
// 03 · educationiiit pune · 2020→2024

Where I learned.

The degree taught me fundamentals. The internships and side projects taught me to ship.

degreeBachelor of Technology · Computer Science and Engineering
schoolIndian Institute of Information Technology, Pune (IIIT Pune)
duration2020 → 2024
gpa8.14 / 10
likedML · transformers · web dev
// 04 · nowapr · 2026

What I'm on.

Inspired by /now pages. What has my attention right now.

buildingAgents that don't break in prod. The stuff from Turgon's browser platform — conservative state writes, pause-and-resume MFA, the retry layer that stops a flaky network from corrupting a good session.
watchingMCP quietly winning — 10k+ servers since launch, OpenAI and Google shipping support. Opus 4.6 landing a Compaction API right as half the field finished building their own context management. LangGraph vs 'just call the model directly' — the line moves every release.
readingMem0 on the state of agent memory. LangChain's context engineering posts. The 'harness engineering is the new discipline' argument — half true, half rebranding.
obsessedWhere the line sits between 'framework needed' and 'model can just handle it.' Sonnet 4.6 moved it again last month. It'll move again next month.
fuelCoffee, coke, 2am ships, the green tick on the last CI run.

The next five years? Either it's the companies that figure out a real production harness — context, memory, guardrails, the unsexy stuff — or Anthropic ships a model that makes every harness you built look like bloated wrapper code.

Which one are you betting on?

I'm just betting on what I can control.