Blog

Projects

  • FlowZero: RL-powered Flow Free solver built on Monte-Carlo Graph Search with expert iteration.
    • Also includes a stochastic puzzle generator and a SAT-based oracle for stress-testing & benchmarking.
  • DreamPy: Python rewrite of the Dreamcoder[*] framework.
  • ThrowdownTV: Low-latency RTMP live video streaming platform (React, Express, Rust, Typescript).
  • Formify: Computer vision tool to convert photos of tables into accurate CSV files.

Education

  • Berkeley Seal University of California, Berkeley 2023—2027
    • Bachelor's of Science, Electrical Engineering and Computer Science.
    • 3.77 GPA (Eta Kappa Nu Honor Society).
  • The Residency Logo The Residency Summer 2024
    • Startup incubator program for ambitious students.
    • Stayed at Arcadia House in Berkeley, CA.
  • Stanford University Stanford University Summer 2022
    • Stanford Summer Session.
    • 4.11 GPA.
  • Moorpark College Seal Moorpark College Dual Enrollment, 2019—2023
    • Honors (Top 25% GPA).

Research

Publications

Left: agents deliberating around a round table. Right: candidate constitutions ranked on a leaderboard with the fittest one selected.

Internal vs. External: Comparing Deliberation and Evolution for Multi-Agent Constitutional Design

Hershraj Niranjani, Ujwal Kumar, Phan Xuan Tan

ICML AIWILD Workshop, 2026

Should the behavioral norms governing multi-agent AI systems emerge from within, through agent self-governance, or be discovered from outside, through optimization? We run the first controlled comparison of internal deliberation and external evolution across three social environments and 180 simulation runs. Evolution significantly outperforms deliberation in collective-action settings, yet its advantage inverts once incentives shift, at times forcing value-destroying cooperation. Tellingly, deliberation never proposes punishment, the cooperation-sustaining mechanism evolution reliably finds, suggesting external optimization wins on peaks while internal self-governance trades peaks for structural responsiveness.

📄 arXiv 📑 PDF
Constitutional Evolution framework diagram showing iterative optimization loop

Evolving Interpretable Constitutions for Multi-Agent Coordination

Ujwal Kumar, Alice Saito, Hershraj Niranjani, Rayan Yessou, Phan Xuan Tan

International Conference on Machine Learning (ICML), 2026

We present Constitutional Evolution, a framework for automatically discovering behavioral norms in multi-agent LLM systems. Using LLM-driven genetic programming, we evolve constitutions maximizing social welfare. The evolved constitution achieves 123% higher societal stability than human-designed baselines and discovers that minimizing communication outperforms verbose coordination.

📄 arXiv 📑 PDF

Research Experience

  • LLM-guided program synthesis at BAIR[*] under Matteo Guarrera[*] & Carlo Bosio[*].
    • Won the EECS Evergreen Undergraduate Research Award for this work.
  • Computationally verifying Berry-Esseen rates for sparse random d-regular graphs.[*]
    • In collaboration with Leo Nagel[*].
  • Multihorizon timeseries forecasting via deep learning and attention.
  • Physics-informed neural networks for PDE solvers.
  • Neural circuits & decision-making at the Wilbrecht Lab[*] under Albert Qu[*].

Teaching

Course Staff

  • UCS1 (Tutor): CS 170 (Fall 2025)
  • Academic Intern: CS 61A (Spring 2024)

Computer Science Mentors (CSM)

  • Senior Mentor: CS 70 (Fall 2025)
  • Junior Mentor: CS 61A, CS 70 (Spring 2025)
  • 5.0/5.0 average student rating.

Coursework

Fall 2025

(Enrolled, withdrew mid-semester)

  • EECS 229A: Information Theory and Coding (Graduate)
  • COMPSCI 270: Combinatorial Algorithms (Graduate)
  • COMPSCI 189: Machine Learning
  • COMPSCI 370: Teaching Computer Science

Spring 2025

  • EECS 126: Probability and Random Processes
  • PHYSICS 137A: Quantum Mechanics
  • COMPSCI 61C: Great Ideas of Computer Architecture (Machine Structures)
  • COMPSCI 198: Directed Group Studies for Advanced Undergraduates

Fall 2024

  • COMPSCI 170: Efficient Algorithms and Intractable Problems
  • EECS 127/227AT: Optimization Models in Engineering (Convex Optimization)
  • PHILOS R1B: Reading and Composition Through Philosophy

Summer 2024

  • COMPSCI 70: Discrete Mathematics and Probability Theory

Spring 2024

  • COMPSCI 61B: Data Structures
  • COMPSCI 197: Field Study
  • EECS 16B: Designing Information Devices and Systems II
  • PHYSICS 7B: Thermodynamics, Electricity, and Magnetism

Fall 2023

  • COMPSCI 61A: The Structure and Interpretation of Computer Programs
  • COMPSCI 195: Social Implications of Computer Technology
  • EECS 16A: Designing Information Devices and Systems I
  • SOCIOL 110: Organizations and Social Institutions

Moorpark College Dual Enrollment (2019—2023)

  • MATH M21: Discrete Mathematics
  • CNSE M84: Ethical Hacking
  • MATH M25C: Multivariable Calculus
  • PHYS M20A: Mechanics of Solids and Fluids
  • PHYS M20B: Thermodynamics, Electricity, and Magnetism
  • PHYS M20C: Light Optics, Wave Motion, and Modern Physics
  • CS M125: Programming Concepts and Methodology

Industry

  • Member of Technical Staff — Stealth Startup Jan 2026—Present, San Francisco
    • Working alongside world-class researchers and engineers as the youngest member of the AI/ML team.
  • Founding Software Engineer — Virio Sep—Nov 2025, San Francisco
    • Designed and implemented graph algorithms for mining and analyzing social network data at scale.
    • Developed ML and multi-agent systems to identify quality content and automate deep research for scalable content creation.
  • Machine Learning Engineering Intern — Property Finder May—Aug 2025, Dubai
    • Built and deployed deep learning models for forecasting Dubai real estate prices across 5M+ listings using a custom transformer architecture; achieved 9% MAPE over a 6-month horizon.
    • Won “Best Tech” at the 2025 internal hackathon for an Mixture-of-Experts ensemble with 5% MAPE on listing price predictions.
    • Explored interpretability via Kalman filters and physics-inspired modeling (SINDy).
  • Software Engineering Intern — LexARI Aug—Oct 2023, Berkeley
    • Developed core infrastructure for “ChatGPT for legal documents” using retrieval-augmented generation (RAG) on court cases and contracts.
    • Integrated PGVector and built semantic search pipelines to boost document relevance and Q&A performance.
    • Created robust scraping tools and containerized testing infrastructure using Docker; deployed services to Azure.

I will note that I have done all kinds of on-and-off paid consulting work for startups, small businesses, and other organizations, but I won't bother listing those here.

Personal

A few non-work things I like — some photos, a couple random facts, and whatever else I feel like adding.

Photos

Fun facts

Birthday
August 5th, 2005
MBTI
INTJ-T
Favorite Movie
The Dark Knight
Favorite TV Show
The Wire
Favorite Anime
Steins;Gate
Favorite Cuisine
Thai
Favorite Musical Artist
Death Grips
Favorite Algorithm
Dijkstra's Algorithm
Favorite Programming Language
Python
Favorite Course
EECS 126
Favorite Sport
Cricket