building at FutureAGI

I build AI systems, mostly where models meet production — and I make the backend hold.

These days that's mostly Python, Django, Temporal and WebRTC, pointed at voice infrastructure and evaluation systems — workflows, telephony, latency, and the unglamorous plumbing that keeps AI products reliable.

Previously I helped take an AI interview product to 50,000+ users and moved 5M+ production records to ClickHouse. Away from work, I build tools, contribute upstream, and study the systems below the framework layer.

github activity

open profile

experience

full résumé

Feb 2026 — Present · India

FutureAGI

Software Engineering Intern

Building the systems used to test, replay, and evaluate production AI agents across voice and chat.

200+concurrent calls tested

real → replayvoice evaluation loop

own infrachat simulation path

Built and owned Voice Observe to Simulate, turning real production voice calls into replayable simulation scenarios with baseline-versus-replay comparison.

Built the LiveKit and WebRTC audio bridge with Temporal activities, SIP and Twilio paths, credential routing, and call monitoring; tested at 200+ concurrent calls.

Migrated chat simulation off a third-party VAPI loop onto a provider-agnostic LiteLLM-backed engine, taking the critical conversation path onto company infrastructure.

Restored call-recording ingestion after a vendor auth change through authenticated downloads, re-hosting, and resumable historical backfill; also shipped evaluation, RBAC, and security hardening.

Jul 2025 — Jan 2026 · India

Grapevine · Round1 AI

Software Engineering Intern

Owned product and platform work for AI interviews, high-volume data, and human-in-the-loop moderation.

60 dayszero to launch

10×faster analytics

1K+ / dayposts moderated

Built ReadyAI from zero to scale in 60 days, growing the AI mock-interview product to 50,000+ users with a two-engineer team.

Designed real-time audio, automated evaluation, feedback, and AI-versus-AI content-generation systems.

Migrated more than 5 million production records from MongoDB to ClickHouse with minimal downtime, unlocking 10× faster analytics.

Built a moderator dashboard processing 1,000+ posts per day through LLM-based auto-moderation and human review workflows for 20+ moderators.

selected work

all projects

no.01 live

HackScrapped · 2025

HackScrapped

A discovery product that turns fragmented hackathon listings into one searchable, continuously refreshed catalogue.

12K+users in 30 days

500+monthly active users

#1Peerlist · week 9

why it matters

Reached 12,000+ users in 30 days, sustained 500+ monthly active users, and ranked #1 on Peerlist in week 9 of 2025.

under the hood

Cheerio handles lightweight pages, Puppeteer handles JavaScript-heavy sources, scheduled jobs refresh every six hours, and Redis avoids repeated scraping and database work.

Next.jsPuppeteerCheerioRedisCron

Live site

no.02 live

XpressPrints · 2025

XpressPrints

A production commerce and operations platform built for a real printing business, not a storefront demo.

liveoperating business

Razorpaypayment workflow

Shiprocketorder fulfilment

why it matters

Deployed for an operating business and actively onboarding paying customers.

under the hood

One operational surface covers the storefront, automated order processing, business analytics, Razorpay payments, and Shiprocket fulfilment.

Next.jsRazorpayShiprocketAnalytics

Live site

no.03 live

Placement Predictor · 2025

Placement Predictor

A placement-readiness system that combines academic, coding-platform, GitHub, skills, and experience signals.

600+students

4 yearsplacement data

500+companies scored

why it matters

Used by 600+ students, backed by four years of placement data and company-specific scoring across 500+ companies.

under the hood

Produces eligibility, assessment, interview, final-placement, and package estimates with AI-assisted analysis, rule-based fallbacks, improvement guidance, and shareable result cards.

Next.jsAI scoringData pipelinesAnalytics

Live site

no.04 source

$ ./proxy --port 3000

clientsthread poolLRU cache

listening · concurrent · cached

C systems · 2024

Proxy Server

A multithreaded HTTP proxy with POSIX threads, socket-level networking, semaphores, and an LRU cache for concurrent clients.

3systems concepts

POSIXthread model

LRUcache policy

why it matters

Combines multithreading, socket networking, synchronization, HTTP parsing, and cache-backed responses in one framework-free implementation.

under the hood

A shared LRU cache reuses repeated resources while semaphores coordinate shared server and cache state across concurrent client threads.

CPOSIXSocketsLRU

source

start here

Let's build software that survives production.

“The most useful engineering starts where the happy path ends.”

Send an email Usually replies within a day

Have an AI systems problem?

Let's trace the workflow, find the bottleneck, and make it reliable.

Building something ambitious?

I can help move it from prototype through the operational details of production.

Prefer the evidence first?

Review the work above, open the résumé, or inspect the source on GitHub.