ARC AGI 3 agentic solver
An agentic framework that reasons through ARC AGI 3 tasks end to end - the first individual submission to pass the benchmark.
Agentic AIReasoningEval
APPLIED AI ENGINEER · ABUNDANT.AI (YC W25)
Building agents that reason, act, and ship.
I'm Pratyush Shukla - friends call me Pratty. I work on agent infrastructure, teleoperation, and agentic workflows, with a research bent toward reinforcement learning, computer vision, and SWE automation. I like systems that close the loop between reasoning and the real world.
// FOCUS
What I spend my time thinking about
Agentic AI
Planning, tool use, and long-horizon agent workflows that hold up outside the demo.
Reinforcement Learning
Reward shaping and policy work for sequential decision-making and control.
Computer Vision
Perception for robotics - detection, pose, and teleoperation pipelines.
SWE Automation
Agents that read, write, and verify code as part of the development loop.
Neuro-fuzzy Systems
Hybrid learning and reasoning where crisp rules and learned signals meet.
Sonomorphic Computing
Exploring sound-shaped computation as an unconventional substrate for inference.
// SIGNALS
1stindividual
Passed the ARC AGI 3 reasoning benchmark
First individual to clear ARC AGI 3 using an agentic framework - a benchmark built specifically to resist memorization and reward genuine reasoning.
18K+downloads
DeFiLlama-Curl on PyPI
Creator and core developer of an unofficial Python 3 wrapper for the DeFiLlama API, now past 18,000 downloads and used across DeFi tooling.
View on PyPIICLR2026
Reviewer, RSI Workshop
Volunteer reviewer for the Reasoning and Symbolic Intelligence (RSI) Workshop at ICLR 2026, reviewing submissions at the reasoning frontier.
// SELECTED WORK
DeFiLlama-Curl
Unofficial Python 3 wrapper for the DeFiLlama API. Typed, batteries-included, and past 18K downloads on PyPI.
PythonOpen SourceDeFi
VisitTeleoperation & agent infra
Infrastructure connecting human teleoperation to autonomous agents at Abundant.ai - the loop between demonstration and policy.
RoboticsInfraAgents
Vision pipelines for robotics
Detection, pose, and perception systems built at Orangewood Labs to drive real manipulators in the physical world.
Computer VisionRoboticsPerception
// NOW & RESEARCHWhere the attention is going
NOW
Applied AI Engineer at Abundant.ai (YC W25) - agent infrastructure, teleoperation, and agentic workflows.
2026
Volunteer reviewer for the RSI Workshop at ICLR 2026, reviewing work on reasoning and symbolic intelligence.
ONGOING
Exploring sonomorphic computing and neuro-fuzzy systems as routes to more robust, interpretable agents.
// EXPERIENCE
Applied AI Engineer · Abundant.ai
YC W25 · Agent infrastructure, teleoperation, agentic workflows
AI & Computer Vision Engineer · Orangewood Labs
YC W18 · Perception and vision pipelines for robotics
// EDUCATION
MS, Electrical & Computer Engineering
NYU Tandon School of Engineering
Machine learning, systems, and computer vision