Core AI research & engineering
Representation learning, model behavior, evaluation, and the infrastructure needed to turn promising ideas into reproducible systems.
AI SYSTEMS ARCHITECT · 2× ACQUIRED FOUNDER · IIT DELHI CS
I build intelligence
for digital and
physical worlds.
My work has two equal tracks: open-source world models that learn predictive representations, and enterprise AI agents that complete consequential work reliably.
● OPEN TOFounding AI EngineerResearch EngineerStaff AI Systems rolesDelhi NCR · open to SF / global relocation
SCROLL TO EXPLORE ↓
2×Acquired founder
90%Document cases automated
446KMinecraft training windows
20KGolden enterprise cases evaluated
[ TWO TRACKS · ONE OPERATING STYLE ]
01 / OPEN SOURCE
World Models
JEPA systems across images, video, ECG signals, and action-conditioned Minecraft worlds. Built from first principles and tested with probes, retrieval, collapse diagnostics, controlled ablations, and counterfactual actions.
91.2%recorded actions beat mismatched actions at H=10
02 / PRODUCTION
Enterprise AI Agent Systems
Multimodal agents, governed data foundations, evaluation infrastructure, and human-review systems built for high-stakes operational work—not demo environments.
90%straight-through automation across document cases
[ OPERATING THESIS ]
Build from first principles.
Test what was learned.
Ship what survives.
The same discipline connects both tracks: instrument the system, test the real capability, publish the limitations, and keep humans in the loop where uncertainty matters.
[ RESEARCH & BUILD INTERESTS ]
Where I want to push the frontier.
Deep research, grounded in systems that create real-world value.
Healthcare AI
Learning from physiological signals and clinical data—especially systems that improve access and decisions while respecting the precision healthcare demands.
World models
Models that learn how environments evolve from video and multimodal experience, with a focus on predictive representations, action understanding, and planning.
Agents that create real value
AI agents that complete consequential work end to end, with tools, memory, evaluation, human judgment, and clear measures of economic impact.
[ FEATURED SYSTEMS ]
Real workloads. Measured outcomes.
01Latest open-source research
World Models · Action-Conditioned Latent Dynamics
MineJEPA-SWM
An action-conditioned world model trained on MineRL demonstrations to predict future Minecraft states in latent space—and tested with counterfactual actions on entirely held-out episodes.
Read case study
02Production system
Sutra.AI · Multimodal Agents · Human Knowledge Capture
Address Verification Orchestrator
A multi-stage agent system that verifies people and addresses across messy, unseen real-world document sets—with human-level accuracy and 90% straight-through automation.
Read case study
03Active open-source research
World Models · Video Representation Learning
Mini V-JEPA From Scratch
A compact V-JEPA implementation for learning predictive video representations, built from first principles and tested on real action data.
Read case study
04Implemented system
Business Questions → Verified Data Answers
Safe Business Data Agent
Answers business questions using trusted company data, verified query patterns, and a deliberate refusal when the evidence is not strong enough.
Read case study
05Active open-source research
Representation Learning · Healthcare AI
CardioRep
Learning transferable ECG representations from raw 12-lead waveforms with JEPA-style masked prediction.
Read case study
06Shipped platform
Fragmented Company Data → AI-Ready Context
Enterprise Data Foundation for AI
A shared data foundation that connects company databases, files, APIs, and business definitions so analytics and AI agents work from consistent context.
Read case study
[ OPERATING HISTORY ]
Built from zero. Operated at scale.
Architecture, technical direction, and hands-on delivery.
2026—NOW
Sutra.AI
Chief Technology Officer
Leading the technology function across production agents, evaluation infrastructure, project automation, and distributed data systems. Joined through the acquisition / merger of YouData.ai.
2023—2026
YouData.ai
Founder & CTO · Acquired
Built the enterprise data and semantic layer that became a core part of Sutra.AI’s platform.
2021—NOW
String Ventures / AI SV
Venture Partner / Founding Member
Technical diligence and zero-to-one company building, including LVX / LetsVenture, ConvergeFi, and Buildrun engagements.
2021—2023
ScorePlus / Exampur
Founder & CTO
Built learning products serving 4M+ students.
2020—2021
Directly.live
Founder · Acquired
Shipped and sold a live consultation SaaS product while at IIT Delhi.
2019—2020
CRED
Software Engineering Intern
Built Go, GraphQL, parsing, and notification infrastructure.
[ HOW I WORK ]
01
Research with diagnostics
Loss curves are not enough. I test effective rank, retrieval, wrong-target margins, probes, and controlled ablations.
02
Agency with gates
LLMs reason and explain. Deterministic systems own scores, permissions, execution gates, and critical state transitions.
03
Architecture with evidence
I document what shipped, what was benchmarked, what failed, and what remains designed—not yet implemented.