Built, as a freelancer for a coaching institute, an exam-paper generator now used by 1,200+ students: LLM-written questions go through a Dockerised LaTeX service that outputs print-ready English or Gujarati papers.
Purv Kabaria
Founding Engineer · Backend & Applied AI
- Location
- Surat, India
- Studying
- CS @ NIT Surat
Final-year CS student at NIT Surat who likes being early. I was a founding engineer intern at Career Passport, where I built the resume-parsing queue and the LLM prompt setup, and as a freelancer I built PaperForge, an exam-paper generator now used by 1,200+ students. I own features end to end, from schema to API to deployment.
Currently exploring Data Science, Java and Deep Learning.
Experience
- PaperForge
Freelance Full-Stack Developer
Jun 2025 - Jun 2026
- Delivered an exam-paper generator for a coaching institute, now used by 1,200+ students: pick a syllabus, difficulty and language, get a print-ready paper with an answer key.
- Added LLM question generation that can switch to local models, and cached admin-dashboard reads in Upstash Redis so repeat views skip the database.
- Built the LaTeX-to-PDF service as a Docker image that compiles the generated LaTeX in isolation, for English and Gujarati papers.
Career PassportFounding Engineer Intern
May 2026 - Jul 2026
- Kept API responses fast during bulk uploads by moving resume parsing off the request path onto a BullMQ + Redis queue with 30 concurrent workers (12s median, 30s p99 per resume).
- Prevented duplicate parsed records when workers crashed mid-job, using a content-hash idempotency key and a one-result-per-resume database constraint.
- Let the team edit 20+ LLM prompts without a redeploy by moving them into Langfuse, with the last good version kept in Postgres as a fallback.
- Replaced manual job-form setup with one-click import APIs that map Google Forms and other hiring sites into our schema (8 field types).
Mercurium AnalyticsSoftware Development Intern
Jul 2025 - Aug 2025
- Took engineers out of routine site edits by building an admin-gated CMS (Django, PostgreSQL) so non-engineers edit layout and content.
- Replaced manual spot-checks with an automated regression run that re-tests document extraction on a held-out evaluation set after every model change, before rollout.
- Evaluated the in-house vision model for ID-document extraction on a hand-labeled dataset (passports, national IDs), logging per-category errors to guide fixes.
- Redesigned the internal admin dashboard's navigation and layout around the team's daily content and account-management workflows.
Leadership & Achievements
- Top 10 Finalist, WWT Unravel 2026 (World Wide Technology), solo entry; finalists earned Pre-Placement Interviews
- Rank 94, Amazon ML Challenge 2026, a 72-hour national ML hackathon with 19,900+ registrations
- Winner, Adani Powermind Hackathon, with a day-ahead energy-demand forecasting platform
- Top 15, EEG/EMG Foundation Challenge 2026 (NeurIPS 2026 Brain & Body Workshop), solo entry
- 1st Runner-up, WebOsmotic AI/ML Hackathon (20 shortlisted on-site)
- Top 100 nationally, ZS Tech Challenge
- Finalist at 12 national-level hackathons
- Chairperson, Cutting Edge Visionaries (Fintech Club), SVNIT
- Developer, ACM SVNIT Student Chapter
Tech Stack
React
Next.js
TypeScript
JavaScript
HTML
CSS
Tailwind CSS
Bootstrap
Vue.js
Three.js
Vite
Expo
Competitive Programming
Featured Projects
All projects
Self-hosted RAG over PDFs, scans, audio and video in 5 independently scalable services. Retrieval benchmarked stage by stage on BEIR SciFact (nDCG@10 0.706 to 0.724), duplicate chunks cut from 17% of top-5 results to 0%, and crash-safe Redis Streams workers.
Writing
All postsAmazon ML Challenge 2026
Rank 94Rank 94 at the Amazon ML Challenge, or: how not to compare 6.7 trillion pairs
Ten million messy business records, 72 hours, and a metric that punishes overconfidence. What actually moved the score, and what was merely fashionable.
Adani Powermind Hackathon
WinnerWe won the Adani Powermind Hackathon. My model was 4x worse than I thought.
Our team won with an energy-forecasting platform. I built the forecaster, reported a dazzling 0.65% error, and later discovered the dazzle was a leak. Here is the autopsy, and the honest number.
WebOsmotic AI/ML Hackathon
1st runner-up1st runner-up at WebOsmotic: round one banned AI, round two required it
Twenty people shortlisted from resumes, then two rounds with opposite rules: build a RAG API by hand in two hours, then fix a broken production codebase with an AI agent riding shotgun.
Get in Touch
Reach out
I'm always up for collaborating on interesting projects or just chatting about tech. My inbox is open, reach out via your favourite platform!
- purvkabaria@gmail.comQuick questions & inquiries
- Connect on LinkedInRoles, referrals & networking
- Follow on GitHubCode, side projects & commits
Usually replies within a day
Open to internships & full-time roles
Send a message
Prefer to write? Drop a note here and it lands straight in my inbox.