Available nowBengaluru or remote

Forward deployed engineer for AI agents

I'm Prem Kumar S. Five years shipping web platforms, the last 18 months embedded at Quince, where I built SprintCTL to put coding agents on real sprint tickets.

SprintCTL 200–500 tickets a week 2–4h → 15–30m to production Still running after I left

SprintCTL. Coding agents I built alone at Quince.

  • 200–500Tickets a week through agents
  • 2–4h → 15–30mFrom issue to a fix in production
  • ~70%Done end to end, rest caught in review
  • ~10Engineers, still using it after I left

One graph of the org

Jira, Slack, GitHub, Confluence, Google Docs and Datadog linked into JEV, so each ticket arrives with its code, owners and threads.

Any coding agent

Claude Code, Cursor and Codex behind one orchestrator, running in the engineer's own checkout through a local bridge.

Asks the right engineer

A stuck agent asks on Slack. It picked the wrong people at first and got accurate as the graph grew.

Humans own the merge

Agents work up to the pull request, never past it. Every change needs two human approvals.

Black Friday on the Quince storefront. I pre-scaled SSR and the BFF for 4–5x traffic and held p95 under 70ms.

Visit quince.com
Quince's storefront: search, category navigation and the holiday gift list hero, the pages served by the SSR platform

What I built inside each Quince request. Embedded at Quince by 91Social for 18 months, April 2025 to September 2026.

4–5x traffic

Pre-scaled SSR and BFF for Black Friday

p95 held under 70ms for SSR and 50ms for the BFF at peak.

500K+ a day

Moved pages to per-request SSR

Collection and product pages off the CDN, plus the first personalized “Picked for you” path.

0 flicker

Ran Statsig experiments on the server

The variant renders into the HTML and the browser SDK starts with the same values.

2.1MB → 720KB

Fixed a 2x LCP regression

The mobile carousel now loads off-screen slides only after the first touch.

Every PR

Hardened the Playwright pixel tests

A pixel passes only if the vendor's server answers, and crashed runs no longer report to Slack as passing.

6 ad platforms

Wired server-side conversions

Click IDs for Reddit, Snapchat, X, Pinterest, AppLovin and OpenAI; Reddit CAPI live before BFCM with $10K in credits.

A network interceptor for pixel and conversion calls. I built it alone at Quince, so a failing pixel, CAPI or sGTM call raises an alert fast.

  • Hybrid capture. Monkey patches on fetch, XHR and sendBeacon, plus a service worker for the requests patches can't see.
  • Watches the money calls. Tracking pixels, Conversions API calls and server-side GTM, the calls the ad budget depends on.
  • Alerts on failure. A vendor that stops answering is flagged quickly, instead of surfacing later in ad reporting.
  • Lightweight. About 13µs per request and a 3.6KB bootstrap in benchmarks.

Agents that run around the clock. Self-hosted on a KVM 2 VPS with OpenClaw and Hermes Agent.

Work arrives by chat or by schedule

Each agent gets a job, the tools it needs and a place to report back.

  • Runs unattended. Scheduled jobs and long tasks keep going while I'm away from the laptop.
  • Uses real tools. Browser, shell and APIs, scoped per agent, on a server I patch and monitor.
  • Answers in chat. Results, drafts and alerts come back as messages I can act on.

Products I ship on my own. Live apps, one in development, three npm packages.

Private beta · waitlist open

Rovelle, cooking videos into recipes

Paste an Instagram, TikTok or YouTube link and get steps you can cook from, read aloud while you cook.

Join the Rovelle beta
rovelle.app: turn any cooking video into a recipe you can actually cook from, beside a phone parsing an Instagram link

Live · npm 1.0.0 · MIT

Recall, a Claude Code skill for career evidence

Rebuilds what you shipped from git history into a cited brag doc, résumé, role fit and interview prep. Local-first, no servers.

See Recall
Recall's site: turn your git history into interview-ready proof of impact, beside a rendered engineering retrospective

In development

Tare, tracing for AI execution

Every step an agent takes, with its tokens, latency and cost, so teams can see what slows them down and what it costs.

Join the Tare list
Tare's landing page: profiling, tracing, optimization and latency panels around the Tare wordmark

Live · npm

Agent skills on npm

Small tools for people who work with Claude Code every day. About 2,400 downloads across my three packages this year.

  • brutal-review18 adversarial review skills for Claude Code
    1.5.0
  • talk-to-me-in-htmlTurns any topic into a one-file interactive HTML explainer
    1.2.0
  • @premdevai/recallEvidence-backed career records from git
    1.0.0
See my npm packages

Rovelle turns cooking reels into recipes. My own product, from the pipeline to the apps.

See Rovelle
Rovelle's product design: brand, recipe card, cook mode timer, cookbook, grocery list, meal planner, iOS and Mac apps

One reel in, a structured recipe out

A TypeScript job pipeline with retries and spend budgets on every model call.

  • Backend. Fastify, Postgres and BullMQ workers
  • Speech. Sarvam Saaras and Whisper
  • Vision. OCR on sampled video frames
  • Structure. Gemini turns it all into recipe JSON
  • Client. React Native app

Smaller things I've shipped. Answers that cite the rule they used, and a data story, built for Tamil Nadu.

10 Porutham: a jathagam chart beside the matching result, 5.5 of 10 points, with rajju and dosham notes in Tamil

Live · Tamil and English

10 Porutham, Tamil marriage matching

All ten poruthams and doshams, follow-up questions in Tamil or English, and every answer shows the rule it used.

Try 10porutham.in
NILAM scheme matcher: sector, project size and promoter profile on the left, a computed ₹4.4L support estimate with cited schemes on the right

Live · Tamil and English

NILAM, a Tamil Nadu scheme matcher

Matches an enterprise profile against cited, versioned government scheme rules and TANSIDCO industrial plots.

Try NILAM
Parandur Files: the land rush before the runway, with a map and land records

Live · data story

Parandur Files, a land-rush data story

Twenty years of land prices and speculation around Tamil Nadu's proposed, now scrapped, Parandur airport.

Read Parandur Files

Forward deployed work I've already done. 18 months embedded at a client: their stack, their traffic, their vendors on the other end.

MetaRedditPinterestOpenAI

Integrates other companies' APIs

Fixed Meta CAPI deduplication across storefront and checkout, and routed credentials per storefront.

DatadogKubernetesArgo CDOpsgenie

Owns production at peak

On call through sales events: resolved 5xx spikes and shipped hotfixes with Datadog and Argo CD.

ReactNode.jsPostgreSQLPlaywright

Ships the whole path

React and React Native, Node back ends, Playwright in CI. Led a small team rebuilding Zillit's web app.

Tools I've used in production

  • React
  • Next.js
  • TypeScript
  • Node.js
  • NestJS
  • Playwright
  • Kubernetes
  • Argo CD
  • AWS
  • Sentry
  • Postgres
  • Redis
  • Gemini
  • OpenAI
  • Docker
  • Linux

Experience at four companies. Since December 2020, after a B.E. in Computer Science.

  • SDE-3, Storefront Platform and Martech 91Social, embedded full-time at client Quince · Bengaluru
    Apr 2025 – Sep 2026

    SSR, BFF scaling, experiments, conversion APIs and peak-season on-call. Built SprintCTL alone: about 10 engineers, 200–500 tickets a week, fixes to production in 15–30 minutes instead of 2–4 hours.

  • SDE-II Zillit · Delhi
    Nov 2023 – Dec 2024

    Rebuilt the web app with React and Redux: encrypted chat, calls and push, used by BBC, Netflix and Prime.

  • React JS Developer Antino Labs · Gurugram
    Jul 2022 – Jun 2023

    Migrated LIC Mutual Fund (KFintech) from .NET to React and Node.

  • Frontend Web Developer Treatment Traveller · Noida
    Dec 2020 – Jul 2022

    Built Hosplan in React: search hospitals and doctors across India and book appointments.