Enterprise work here is described by what it does. Clients and employer stay private; what I build independently, I'll name. Read it less as a portfolio than as one repeatable move: hand me a hard, expensive problem and it comes back live, accurate, and handed off clean.
The shape of the impact is consistent: thousands of human hours automated, six-figure ($) savings in the first months, and value that compounds into the millions as the system runs. These are the ones I can describe — there's more that isn't public yet.
Extracts the clauses that matter, flags the risk, and runs the review work that used to be done by hand, across tens of thousands of documents.
Saving six figures ($) in costs from its first months, and still compounding.
Turns the messiest real-world documents into clean, structured data. Multi-tenant, live, and accurate enough to trust without a human in the loop.
Treated by leadership as an emerging revenue line.
Social-media management across platforms and languages, with a real media pipeline at its core: transcription, translation, and voice dubbing. My own venture, built end to end. Brief on purpose.
Built & run as founderRetrieval over a library of historical proposals plus semantic search, so responses to complex requests get drafted in context, grounded in prior work instead of a blank page.
A streaming, staged-concurrency redesign of a document-processing path: download, CPU work, and API calls pipelined instead of run in series. About 30–40× the throughput of what it replaced, on the same hardware.
Yantra: my own AI development environment, built from scratch to keep agents, projects, and context coherent across dozens of parallel workstreams.
A full, signed, notarised app: 600+ commits, 900+ tests, my daily driver for months. Not a plugin, not a fork. Other developers have asked for early access.
Stenograph: a private knowledge engine. Deep research grounded in your own context and projects, with an MCP layer that hands the answers straight to the agents I already run.
On track to process well over a million hours of media by late 2026. Research as callable infrastructure, not a portal you visit. Built solo.
A telephony agent that holds a live conversation: streaming speech-to-text into an LLM and back to voice, with voicemail detection and barge-in. It handles being interrupted, and it answers in under half a second.
Self-funded research into understanding media at scale: ingesting tens of thousands of items, running vision-language models across my own GPU fleet, and benchmarking models head-to-head to find what actually works. The hardware involved is the kind most engineers only touch through an employer.
Want the detail behind any of these? Ask.
This site is my own work, end to end — like everything it describes. · v1.3.0 · cf9ec25
Hey, good to meet you. Quick heads up: I'm an AI version of Rahul, trained on his work, so I can talk through what I build and how I work any time. What brings you by? Got something you're trying to get built?