Case studies

Three pieces of work, each told through the problem, what was built and the outcome. Status labels show what shipped in my job and what is a prototype or a research experiment.

How to read the status labels

Built and launched at PPFAS: Work I built at PPFAS that is launched or in production use.
Prototype · not deployed: Software I built that is not deployed and has no automated tests or users, so there is no external impact to report.
Research prototype: A feasibility experiment, not deployed; results are technical measurements, not validated with the people it is meant to serve.

Built and launched at PPFAS

Building public websites, portals and internal tools at PPFAS

At PPFAS I built public websites, the Gift investment platform and internal portals that are in production use, with one combined fixed-cost saving of ₹20+ lakh across seven builds.

Problem: Business teams depended on a separate tech team for website and workflow changes; static pages and management portals reduced that dependency.
What was built: gift.ppfas.com, the Gift investment platform (integrated with a third-party payment gateway, with international payments), wealth.ppfas.com, a management portal for amc.ppfas.com, an event-management system and a Career portal.
Outcome: ₹20+ lakh in fixed costs, combined across seven builds, and ₹10+ lakh in annual recurring cost, attributed jointly to the event-management system and the Career portal.

Prototype · not deployed

SYNAPSE: an AI SaaS product prototype for LinkedIn personal branding

A full-stack prototype for deciding what to post about on LinkedIn and drafting it in the user’s voice; nothing is posted without the user’s explicit action.

Problem (premise, untested with users): Posting on LinkedIn consistently takes hours of content work, and a general chatbot starts from an empty prompt.
What was built: Trend discovery from three sources, three variations of each post, and a style profile learned from the user’s own edits.
Outcome: No external impact — prototype, not deployed, no users.

Research prototype

SignBridge: mapping spoken video to a sequence of Indian Sign Language concepts

A research prototype testing whether spoken video can be transcribed, turned into sign-language concepts and played back as signs in sync with the video.

Problem: Spoken video is not automatically available in sign language, and captions are not the same thing as sign language.
What was built: A pipeline that turns speech into a sequence of signs synced to the video (only 2 of its 209 vocabulary entries are real sign clips; the rest are placeholder cards), plus a separate experiment using motion from real Indian Sign Language (ISL) clips to animate a 3D avatar.
Outcome: Technical findings only, not tested with Deaf users. An arm-reach correction for the avatar cut average wrist-path error by roughly 85% on one HELLO clip versus the earlier approach (0.2405 → 0.0354); a later fix that stopped the arm clipping into the hip raised it to 0.0665. On FRIEND, where the hands touch, the right hand was detected in only 34% of frames (local, uncommitted research).

Earlier projects

Earlier work includes Book Nook (a full-stack book store) and a real-time multiplayer Mafia game. See earlier projects.