I build software that removes unnecessary work.
Backend systems, AI applications, and automation pipelines — for problems that actually exist. An NGO was paying ₹10L/year to digitize farm records by hand; now an OCR pipeline does it at 95% accuracy. Students waste hours applying to dead job postings; now a detector flags them before they do.
B.Tech ICT at Dhirubhai Ambani University. Most of my work runs on Python, FastAPI, PostgreSQL, and n8n, built with Claude Code as an AI pair-programmer — I do the architecture and the evals, and I test like I don't trust either of us. The numbers below are measured, not estimated.
|
Problem — ~1 in 5 internship postings is a ghost: never going to be filled.
Solution — A discovery platform with a 5-signal ghost-job detector and a response-likelihood model trained on real application outcomes.
Impact — Flagged 11% of "live" postings as 100+ days stale; prediction calibration improved Brier 0.219 → 0.181 on a leak-free temporal split.
|
Problem — LLM API bills grow faster than the products they power.
Solution — Self-hosted middleware: 7 composable optimization strategies behind an OpenAI-compatible proxy, with a quality gate that auto-rolls-back degraded prompts.
Impact — 25–45% cost reduction across 9 providers; 27.6% token reduction at 0.945 quality parity on HotpotQA (n=55, bootstrap CIs).
|
|
Problem — Product research means reading 400 reviews and trusting none of them.
Solution — A 12-agent pipeline that aggregates, verifies, and ranks product evidence, with deterministic mention counting and 5-provider LLM failover.
Impact — Research runs that took an evening finish in minutes, with sources attached.
|
Problem — Concurrency is easy to claim and hard to prove.
Solution — A multi-process air-traffic simulator in C: |
- Building — expanding ITOL's provider adapters; instrumenting InternPilot's outcome loop with more real application data
- Learning — distributed systems fundamentals; making evaluation harnesses a default habit, not an afterthought
- Open to — software engineering & AI-automation internships
LinkedIn · Email · Ahmedabad, India
If a task takes ten minutes and happens twice a week, I have already spent eight hours automating it. No regrets.