Re-architected purchase and upgrade flows to launch add-on subscriptions, contributing to ~20% revenue growth; team won “Astral Team Award” (2026)
Co-led the migration of 14M+ JioCinema subscribers to Hotstar in 14 days with zero data loss and zero missed renewals; architected the ingestion on Kafka with an audit log and dead-letter replay
Shipped the S3 plan sync and in-memory plan SDK behind the P0 internal read path (~50K RPS peak); architected and led its rebuild, load-tested at 15 ms P99 on 15–20% fewer pods
Owned the plan service behind 1K+ subscription plans in 25+ markets; automated the Disney+ Hotstar → Disney+ plan migration (6K+ config changes), cutting ~3 hours of manual launch-day work to under 15 minutes including verification; reused by a teammate for the follow-on market handover
On-call through IPL at 30M+ concurrent viewers; led a SEV-1 at JioCinema when Akamai blocked ~200K payment-gateway renewal webhooks, restoring ~70K renewals within 3 hours through an expedited allow-list change and a fix to the reconciliation cron
~20% compute savings and zero production incidents from a Graviton + Java 21 migration, packaged as Claude and Codex agent skills and reused across ~5 legacy Java services; root-caused a Java auth-library issue while onboarding to the org's service-level testing framework initiative, a fix 2–3 other teams then applied; LLD/HLD hiring panelist (20+ interviews)
2024 – 2026
FamPay (YC S19)Software Engineer, Backend
Joined as the second backend engineer; led a team of four on FamCard: 5M+ virtual cards, ₹10M+ daily transactions, 20–30% lower delivery cost
Architected the gift card platform (₹60 Cr+ cumulative GMV) and built recharge and bill-payment products, around 60% of monthly revenue
30+ backend interviews; onboarded new engineers and mentored 5+ interns and juniors into full-time roles
2020 – 2024
InstamojoSoftware Engineer, Backend
Automated merchant onboarding, cutting KYC resubmissions by ~25%
Built reseller product ingestion that onboarded ~15K products in its first month; converted from intern to full-time (PPO)
An 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.
Evals pick the model: every model, prompt or provider change is scored field by field against a labelled golden set before it ships. I evaluated Claude, GPT, Gemini, Qwen, Mistral and Sarvam side by side, and the cheapest models that got every ledger-critical field right (total amount, invoice value and actual weight) went to production.
Phased model rollout: production runs Gemini 3.8 Flash and GPT-6 Luna, switched tenant by tenant from the admin panel with no redeploy. On the latest eval, GPT-6 Luna is 8× cheaper and ~45% faster than Gemini Flash, with those ledger-critical fields at 100%.
Guardrails and accept policy: GSTIN checksum and state checks, date and total reconciliation, and fuzzy station matching. A document is auto-added only when every mandatory field passes; otherwise it's flagged for human review with the exact fields named.
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 and latency.
Model selectionEvalsStructured outputsHuman-in-the-loopLLM cost metering
An AI product that helps Indian consumers draft and file consumer-court complaints without a lawyer. I built and run it end to end.
Vision and tool use: Claude reads uploaded evidence (PDFs and photos) to write court-format annexure descriptions, and uses its web-search tool to find the opposite party's registered address and legal contacts.
Runtime model routing: a different model per task (Sonnet drafts, Haiku handles lookups and vision), configurable per user cohort from the admin panel without a redeploy. Each paid document pins the model plan it was drafted with, so re-renders are reproducible.
Cost and abuse controls: a per-call cost ledger in INR, hourly and daily AI caps per user, and input limits on every AI feature.
LLM observability and legal review: calls, tokens, latency, errors and INR cost tracked per feature, provider and model. Drafts go to a consumer-law advocate for sign-off before public launch.
A multi-tenant gateway, built for a logistics ERP software company, that connects each customer's own WhatsApp Business number to their ERP install. I designed, built and run it on my own; three client businesses are onboarded.
Self-serve onboarding: a customer connects their number through Meta Embedded Signup. The code-to-token exchange happens inside the service, so the app secret never reaches a customer install, and the token is handed over exactly once.
Webhook routing: every customer's Meta webhooks arrive at one endpoint, are signature-verified and routed per account. ERPs receive events over a long-poll in about a second, a per-minute pull, or a signed HTTPS push with back-off; delivery is at least once.
Security: HMAC-signed requests, AES-GCM encryption at rest, an SSRF guard on outbound pushes (public IPs only, no redirects), and TOTP admin sign-in behind an IP allowlist.
Small footprint: one Go binary on SQLite (WAL) with continuous Litestream backup to S3. Load-tested on a laptop at ~8.5K signed webhooks/s (p99 ~12 ms) and ~5 ms p99 delivery across 1,000 concurrent long-polls, with zero errors.
Multi-tenantWebhooksLong-pollingHMACSQLite
Go · chi · SQLite (WAL) + Litestream · Prometheus · AWS Lightsail + S3 via Pulumi
Open source, education and certifications → Background