University of Pennsylvania

I build AI systems for real users.

Math and CS at Penn, building across AI agents, ad tech, LLM research, and 3D worlds.

Jimmy speaking into a microphone while interviewing an AI educator on a podcast
Jimmy flashing a peace sign beside a friend on a tennis court at night
Portrait of Jimmy Zhong

About

I’m Jimmy, a Mathematics and Computer Science student at the University of Pennsylvania. Over the last two years I have built AI products with real users: an iMessage networking agent, an ad-campaign manager used by paying customers, and a climate-risk model used by more than 6,500 people.

I care about the whole stack, from shader code to the interface a customer sees. Away from the keyboard: squash, tennis, guitar, poker, and travel.

Education
University of Pennsylvania, B.A. Mathematics and Computer Science. GPA 3.81.
Research
LLM self-training, coastal climate risk, and environmental economics.
Languages
English, Mandarin (native), Japanese (conversational)
Based in
Philadelphia, PA
Jimmy in a red Penn polo greeting a guest at the check-in desk of an SF Tech Week event
Organizing an SF Tech Week event for a community-built 3D world
Jimmy lying on a bed beside his ragdoll cat
With Bubu, my ragdoll cat

Experience

AI products across four fields: generative 3D worlds, advertising, website building, and networking.

Jimmy at a lectern in a lecture hall, running a hackathon workshop
Running a workshop and judging at a hackathon
Jimmy with his hackathon team, sitting on red stairs
Competing in a hackathon as team lead
Aug 2026 - Oct 2026

Altverse

Software Engineering Intern, Generation Pipeline

Chrona (chrona.world): a persistent shared world whose assets, buildings, and scenes are generated by its users from text prompts.

  • Standardized the generation pipeline into a serialization layer that chains single assets into persistent, editable scenes
  • Shipped it into the live public world and the LA Tech Week showcase, drawing 10 artists with 10K+ followers each
  • Designed and built personalized 3D interactive worlds for artists with Three.js and Tripo assets, reaching 200K+ views on X
Mar 2026 - Sep 2026

Welta AI

Founding Full-Stack Engineer

AI advertising campaign management (welta.ai). Five-engineer founding team, three paying business customers, ten-plus live campaigns.

  • Architected the multi-agent campaign manager: cross-platform performance review went from 20 manual minutes to a 5-second query
  • Built the LLM-driven creative-generation flow for programmatic ad creative, cutting variant generation from over an hour to 2 minutes
  • Built the platform in Next.js and TypeScript with Python LangGraph agents on Supabase Postgres, integrating the Meta and Google Ads APIs
Jun 2026 - Aug 2026

DeepWisdom (MetaGPT)

Software Engineering Intern

Atoms (atoms.dev): an AI website-building agent with 1.2M monthly visits and 70K GitHub stars.

  • Designed and ran A/B tests routing tailored prompts by user intent, lifting click-through on generated sites by 23%
  • Analyzed a 20,000-user experiment showing style guidance made users 15% more likely to publish, then shipped the feature
  • Built the event-tracking robot for the Atoms experiments, so new tests announce themselves and data queries take half the time
Sep 2025 - Mar 2026

Franklink

Co-founder and Technical Lead

The first AI networking agent native to iMessage.

  • Built the agent on LangGraph with a ReAct loop, Redis for cache and Supabase for storage, deployed on AWS
  • Raised the angel round from a Y Combinator China founding team member and set the technical direction for the team
  • Led the weekly discussions that broke the roadmap into per-engineer tasks and marketing targets, and reviewed progress with the team

Selected work

Five things I built end to end: a painted 3D world, a renderer, a game, a research tool, and a funded product.

2026Real-time ink rendering

World of Painting

A first-person walk inside Along the River During the Qingming Festival: Song-dynasty Kaifeng rebuilt as a 3D world that still reads as ink on silk.

  • Real-time ink pipeline: silk ground cut from the original scan, brush-contour outlines, water drawn as strokes
  • 24 human and 5 animal figure recipes across 13 poses, all true geometry with depth and normals
  • 351 tests; every scene judged by an independent critic loop, 13 rounds so far
Three.jsTypeScriptGLSLVitePlaywright
Source on GitHub
The Rainbow Bridge of Kaifeng rendered as ink on silk, seen from the south-east bank
2026Real-time rendering

Neon Street

A rain-soaked cyberpunk block you can walk through, delivered as one 787 KB HTML file. No asset packs, no server, no dependencies.

  • Hand-written WebGL2 forward renderer, one draw call per frame
  • 132k triangles, 135 placed lights, 48 shaded per fragment
  • 5,500 rain instances lit by the same light set
WebGL2GLSLPythonBlender
Source on GitHub
A rain-soaked neon alley at night, wet asphalt reflecting magenta and cyan signage
2026Browser game

Odyssey: Nobody

A six-act, first-person retelling of Odyssey Book IX, from the lotus-eaters to the blinding of Polyphemus, running entirely in the browser.

  • Six Three.js scenes, each built and torn down with its own lighting and collision
  • Every string switches between Chinese and English instantly
  • 86 voice lines, 96 passing unit tests
Three.jsTypeScriptViteVitest
Source on GitHub
Title art for Odyssey: Nobody, a ship at sea beneath a stormy sky
2024 - 2026Climate research

Migracoast

A coastal inundation and climate-risk model for nine U.S. states, released as an open-source research tool.

  • NOAA inundation rasters, IPCC AR6 sea-level projections, and storm-event records
  • Risk-score time series for major cities, projected from 2020 to 2100
  • 6,500+ users at migracoast.com
PythonGISscikit-learnJupyter
Source on GitHub
Generated inundation map showing projected flooding along a coastline
2025 - 2026AI agent, funded

Franklink

The first AI networking agent native to iMessage. It learns what you need and who can help, then makes the introduction.

  • LangGraph orchestrator routing every message through onboarding, recommendation, networking, or chat
  • Matches on need and value: what you are looking for against what others can offer
  • Uses your email for context (a conference, a course, a hackathon) and finds people doing the same
PythonFastAPILangGraphRedisSupabaseAWS
Source on GitHub
Franklink logo and wordmark
500+ introductions made by the agent
$50Kangel funding
13person team
2,000+messages handled

Off the clock

Courts, slopes, strings, and the occasional long trip.

Research, awards, and leadership

Papers, prizes, and the organizations I belong to.

Start-Up in a Weekend Hackathon, 2025

Won the Hybrid Intelligence Track

A career-counseling agent, taken from idea to a shipped product in 48 hours.

2023

Finalist, Model Entrepreneur Competition

Pitched a marine waste collection robot to the judging panel.

Summer 2024

MIT Beaver Works Summer Institute

Remote sensing, GIS, Python, and machine learning: the same tools my coastal research runs on.

Research and papers

  1. 01

    Verifier-Guided Self-Training for Large Language Models

    How imperfect verifiers steer iterative LLM self-training. Proved in a logistic model that two verifiers with identical true- and false-acceptance rates can drive training in opposite directions, then tested it on Qwen3 and Llama-3.2 with a PyTorch and LoRA pipeline.

    Submitted to AISTATS 2027
  2. 02

    Migracoast

    The research behind the tool in Selected work: a flood-risk model trained on 790K FEMA claims with spatially blocked cross-validation, with coastal error 27% below a linear model.

    Independent research · 2024 to 2026
  3. 03

    The Impacts of Income Inequality on China’s Carbon Emissions

    Presented at the International Conference on Economics.

    Conference paper
  4. 04

    Coastal biodiversity under a warming ocean

    Modeled Massachusetts coastal vertebrate biodiversity under warming and acidification scenarios using CMIP5 climate projections.

    Climate model

Leadership and activities

  • Wharton China AssociationSince Aug 2025

    Sponsorship Chair. Company outreach for the annual Penn Wharton China Summit.

  • Penn Venture LabSince Aug 2025

    Member. Building alongside student founders in Penn’s early-stage incubator.

  • Wharton Hedge Fund ClubSince Oct 2025

    Global Macro Board. Researching and presenting macro trade ideas to the coverage group.

  • TEDxYouth at Tabor AcademyApr to Dec 2024

    Founder and organizer. Secured the license, built the committee, and drew 100 attendees.

Toolbox

What I reach for, and where it has shipped.

Languages

  • Java
  • OCaml
  • C++
  • SQL
  • HTML/CSS
  • LaTeX

AI and ML

  • PyTorchLLM self-training
  • TensorFlow
  • scikit-learnMigracoast
  • Hugging Face Transformers
  • LangChain
  • Multi-agent orchestrationWelta
  • RAG
  • LoRA/PEFTLLM self-training

Systems and data

  • RedisFranklink cache
  • SupabaseFranklink, Welta
  • AWSFranklink deploys
  • QGISMigracoast
  • PostgreSQL
  • Docker
  • Linux
  • Git
  • pandas
  • NumPy
  • OAuth
  • A/B testingAtoms

Web and 3D

  • React
  • FastAPIFranklink API
  • WebGLNeon Street
  • GLSLNeon Street, World of Painting
  • BlenderNeon Street props

Let’s talk about what you’re building.

Recruiting, collaborating, or just curious: my inbox is open.