I built the prototype that won a state Higher Education contract.

Product Associate at ConveGenius AI. Before that, AceGrade: 213,000 users on zero paid spend, run solo.

Read the case

Higher Education MIS

The applicant-to-admission system that won a state government contract, a bilingual student app plus a role-aware admin portal.

01

What it had to do

A state Higher Education department ran admissions on fragmented manual workflows: registration, eligibility, document scrutiny, merit, allotment, and fee payment all disconnected.

  • Deterministic application lifecycle: scrutiny, merit, allotment, payment as explicit states, so every applicant always knows exactly where they stand.
  • Merit ranking with documented tie-breakers (score, then DOB, then category, then application number), making the algorithm auditable.
  • Document scrutiny with audit history, because government workflows need a paper trail.
  • Bilingual from day one, not as an afterthought.
02

How fast, and with what

Built the winning prototype in 2 to 3 days with Claude Code and Codex and presented it to state leadership. After it won the contract, led the product-ownership workstream, owning requirements, wireframes, stakeholder alignment, and the handoff into the production build.

03

What happened, and what I do on it now

Prototype won a state government Higher Education contract. Production rollout in progress, with me leading the product-ownership workstream for the portal.

The prototype won the work. The production rollout is a team effort, and I own the product workstream on it.

hp-mis-portal.vercel.app

Sign-in required, so there is no screenshot to show.

Users
Applicants (bilingual EN/HI student mini-app) and department staff (State Admin and College Admin portal).
Stack
Next.js 15, React 19, TypeScript, Tailwind, Framer Motion

The work

Fifteen things I built, and which of them are actually running.

11 are deployed and in use. The rest are prototypes, badged as such.

swiftpoll.vercel.app
SwiftPoll, as deployed
Live

Lightweight polling tool the academic training team runs every training cycle.

ConveGenius
repo-sumit.github.io
SwiftLeads, as deployed
Archived

Offline-first lead capture with audio recording, built in hours for an AI summit.

ConveGenius
fmb-demo.lovable.app

Sign-in required, so there is no screenshot to show.

Live

An offline-first survey mini app that lets field inspectors run school surveys with no connectivity, then sync on reconnect.

ConveGenius

Three real build times, and what the model is not allowed to decide.

No averages. These are the only elapsed times I can actually source.

six hours

Pravasi Setu

A morning request delivered after lunch. The founder presented it in Dubai that week.

a few hours

SwiftLeads

Built the night before an AI summit. Zero data loss across an unstable venue network.

two to three days

Higher Education MIS

Built with Claude Code and Codex, presented to state leadership, and it won the contract.

HR Support Bot

  1. Gemini embeddings
  2. NumPy cosine store
  3. SSE stream

No LangChain and no vector database. Answers stay grounded strictly in the ingested HR policy PDF, across English, Hindi and Hinglish.

SwiftChat

  1. local regex
  2. frontend analytics
  3. Groq intent classifier
  4. pgvector RAG
  5. suggestions

Cheapest path first, expensive paths only when needed. A permission guard re-validates every action even when the model proposes it: the model suggests, the system decides.

Where I chose not to use a model

  • Pravasi Setu

    Shipped a rule-based chatbot instead of an LLM, fitting the six-hour budget without faking capability.

  • HR Support Bot

    Rejected LangChain and a vector database to keep the system simple, cheap and debuggable.

Capabilities

What I reach for, and where each one earned its place.

Every tool below is attached to something in the work above.

Product

Discovery
Field interviews and problem-finding before a line of code.
Roadmapping
Sequencing what ships now vs next vs never.
Prioritization
RICE / impact-effort calls under real constraints.
PRDs
Crisp specs that engineers can build from directly.
Stakeholder Mgmt
Aligning state teams, founders, and engineering.

Analytics

SQL
Slicing product data to answer the real question.
Power BI
Executive dashboards and reporting.
Google Analytics
Web funnels, sources, and behaviour.
Firebase
App event tracking and engagement.

Development

  • React
  • Next.js
  • TypeScript
  • Node.js
  • Python
  • FastAPI
  • PostgreSQL
  • Supabase
  • Tailwind
  • Vercel
  • GitHub

AI Stack

Claude Code
My primary pair for shipping at team speed.
Cursor
AI-native editor for fast iteration.
Codex
Code generation and refactors.
Groq
Low-latency LLM inference (llama-3.3-70b).
Gemini
Embeddings and multimodal reasoning.
RAG
Grounded answers from real documents.
Agent Systems
Tool-using, multi-step AI workflows.
Prompt Engineering
Reliable LLM behaviour by design.

What I looked at, and what I changed

Pocket VSK (Unified Portal)

A flat state average hid which schools were actually underperforming for their level.

Benchmarked every level of the hierarchy against its N+1 peer instead, and made the KPI renderer config-driven so a new indicator ships as config, not code.

AceGrade

Funnel data showed drop-off at the quiz evaluator.

Rewrote it to explain why an answer was wrong rather than only that it was. Repeat usage rose afterwards; there is no dashboard export to prove causation.

BDM Capstone

A year of a family grocery store's sales and inventory, unanalysed.

ABC segmentation and demand forecasting, published as an interactive report rather than reproduced here.

Founder first, then the job.

Each leg of this paid for the next one.

Jun 2025 to Present

Product Associate, ConveGenius AI

  • Full-time conversion from the product internship.
  • Full-stack product execution across AI products and internal tools.
  • PRDs, discovery, prototypes, dashboards, and stakeholder demos.
  • Shipped CG Signal, the FMB Survey Builder and the HR Support Bot into production use.

Sep 2024 to May 2025

Product Intern, ConveGenius AI

  • Ran product research, built prototypes, and wrote the documentation and demo assets that went with them.
  • Worked across state education systems, dashboards, AI workflows, and internal tools.
  • Built demo-ready prototypes on short notice, including one presented to state leadership.

Sep 2023 to 2024

Founder, AceGrade

213K+
Users
30+
Countries
1.23M
Sessions
100%
Organic
  • A free, ad-free exam-prep platform, run solo for a year: product, growth, content, community, analytics, infrastructure.
  • Grew a 10,000+ member learning community alongside the platform.
  • Shipped AceGuide too, a campus navigation app used by 5,000+ students during IIT Madras fests, leading a five-person dev team end to end.
  • Managed product, growth, infra, content, analytics, and community.
  • Self-funded the hosting and ran it without a team.

Open to AI Product, APM, PM1 and Founding PM roles.

Noida, open to Bangalore. The fastest way to reach me is email.

or +91 95464 88385