Hey, I'm Shagun Khemka.
I build backend platforms that stay fast under heavy load, and production AI systems that stay cheap and correct.
Senior Software Engineer at JioHotstar, working on the subscriptions and entitlements platform. Previously FamPay (YC S19) and Instamojo. Based in Bengaluru.
Experience
JioHotstarSenior Software Engineer, Backend
Subscriptions platform for 280M+ users. Architected and own the P0 read path serving entitlement checks to Identity and Playback at ~50K RPS, holding P99 SLAs on ~20% fewer pods. Re-architected purchase and upgrade flows (~20% revenue growth, Astral Team Award) and led the cross-team migration of 14M+ JioCinema subscribers. On-call for IPL at 30M+ concurrent viewers.
2024 – now
FamPay (YC S19)Software Engineer, Backend
Led a team of four on the FamCard platform: 5M+ virtual cards, 200K+ physical deliveries, ₹10M+ in daily transactions, and ~30% lower delivery cost. Built recharge, gift card and bill-payment products for 5M+ monthly transacting users, around 60% of company revenue. Ran hiring and mentored 5+ engineers into full-time roles.
2020 – 2024
InstamojoSoftware Engineer, Backend
Automated merchant onboarding, cutting KYC resubmissions by ~25%, and built the reseller product-ingestion service that onboarded ~15K products in its first month. Converted from intern to full-time.
2019 – 2020
Projects
Document extraction platformin production
privateAn LLM pipeline, built for a logistics ERP software company, that turns scanned, handwritten consignment notes into validated, structured data. I designed, built and run it on my own.
- Model cascade: Claude Haiku extracts every document and matches Sonnet on money and weight fields (100%) at ~60% lower cost. Only fields the review policy flags escalate to Sonnet.
- Evals as a release gate: every model, prompt or provider change is scored field by field against a labelled golden set. Eval-driven output trimming cut Sonnet's cost per document by 52% while accuracy rose from 92.7% to 100%.
- Guardrails and human review: schema validation, GSTIN and date repair, and fuzzy station matching. Low-confidence output goes to a review queue instead of the ERP.
- Gateway, cost and observability: Anthropic, OpenAI, Gemini, Mistral and Sarvam sit behind one interface, and tenants can bring their own keys. Every call is metered per tenant, and dashboards track cost per document, escalation rate and latency.
Model routingEvalsStructured extractionHuman-in-the-loopLLM cost metering
Python · Django · Postgres · Claude / GPT / Gemini / Mistral / Sarvam · Pulumi · Prometheus + Grafana
An AI product that helps Indian consumers draft and file consumer-court complaints without a lawyer. I built and run it end to end.
- Grounded drafting: a guided flow collects the facts, checks jurisdiction and limitation rules, and has the user confirm the facts before the LLM drafts a ready-to-file complaint.
- Runtime model routing: a different model for each task (Sonnet drafts, Haiku handles lookups). Each user cohort can get its own models, switched from the admin panel without a redeploy.
- Cost and abuse controls: a per-call cost ledger in INR, with hourly and daily AI caps per user.
- LLM observability and data residency: tokens, latency and failures tracked by feature and provider, with alerts. A Sarvam option keeps data in India, and lawyers review output quality.
Multi-model routingGroundingCost controlsLLM observabilityData residency
Next.js · FastAPI · Postgres · Claude / GPT / Gemini / Sarvam · AWS via Pulumi · Prometheus + Alertmanager
Watch live streams from 14 Mumbai and Pune pandals on one screen.
- Finds live streams through free RSS feeds and one batched API call, staying inside YouTube's daily quota.
- No server and no build step: it runs entirely at the edge on a Cloudflare Worker, KV and a cron that pauses overnight.
- Built and launched in a few hours, with traffic from India, the US, Belgium and Singapore during the festival.
Cloudflare Workers · KV · YouTube Data API · vanilla JS
Open source, education and certifications → Background