Jonathan Yang

Engineer & Artist

Selected Works

{ Scroll }

Hi, I'm Jonathan. I build AI systems end to end, from the cluster to the interface. On the side, I draw.

Disciplines
  • Systems
  • Product
  • Art
Previously
  • NVIDIA
  • Palantir
  • PepprAI (YC W25)
Now on view
Studying Electrical Engineering at Northwestern and doing robotics research.

In figures

Palantir Meritocracy Fellows selected
22 / 800
faster insurance quoting at Palantir
250%
first-author research papers
2
languages spoken
4

Floor plan

Select a room to go there.

RoomI

NVIDIA

Turning a desk of DGX Sparks into one inference machine.

Highlights

  • Built DGX Spark orchestration with vLLM across ring and switched topologies. A switchless four-Spark ring reached about 90% of a mesh's throughput.
  • Worked with EAGLE-3 speculative decoding to speed up decode on the cluster.
  • Integrated Clay, Databricks and Salesforce into an automated sourcing and outreach pipeline for NVIDIA Inception.
  • Built a portfolio-analysis agent for NVIDIA's venture capital arm.
Sparks in a switchless ring
4
of a mesh's throughput
~90%
  • Four gold DGX Spark units standing side by side on an office desk
    Fig. 1Four DGX Sparks, before cabling
  • Four DGX Sparks cabled into a network switch
    Fig. 2The switched topology
  • An exposed GB10 Grace Blackwell board on display next to a DGX Spark
    Fig. 3Inside: the GB10 Grace Blackwell superchip
  • Four workstation towers on a lab bench, joined by a tangle of network cables
    Fig. 4Lab bench, mid-benchmark
Technical notesTopology, Speculative decoding, Go-to-market toolingOpen 3 sheets
Study
I.1
Subject
Topology
Sheet
1 of 3

Four Sparks, one ring, no switch

Each DGX Spark has two 200GbE ports. Four of them can be cabled straight into a ring, with no switch. Benchmarked with AIPerf and vLLM against a four-node mesh, the ring reached about 90% of the mesh's throughput, a cheaper way to run models too big for one Spark.

200GbE200GbE200GbE200GbESpark 12 × 200GbE portsSpark 22 × 200GbE portsSpark 32 × 200GbE portsSpark 42 × 200GbE portsDirect ringno switch
Ring topology, plan viewEach link: 200GbE, node to neighbour
Study
I.2
Subject
Speculative decoding
Sheet
2 of 3

EAGLE-3: guess ahead, verify in one pass

Decoding one token at a time is limited by memory bandwidth, so most of the GPU's compute sits idle. Speculative decoding puts that compute to work. A draft guesses several tokens ahead, and the full model verifies all of them in one forward pass, keeping every token it agrees with. The output is identical to the full model's, only faster.

EAGLE-3's draft is a lightweight head that reads the target model's own hidden states from several layers, not a separate model. It's trained on its own earlier guesses, which keeps acceptance high deep into the draft.

Study
I.3
Subject
Go-to-market tooling
Sheet
3 of 3

Sourcing and outreach for NVIDIA Inception

Inception is NVIDIA's program for startups. I connected Clay for enrichment, Databricks for data and scoring, and Salesforce as the system of record into one automated pipeline. It sources promising companies and kicks off outreach without the manual copy-paste between tools.

  1. 1ClayEnrichment
  2. 2DatabricksData and scoring
  3. 3SalesforceSystem of record
  4. 4OutreachKicked off automatically
One automated pipeline, no copy-paste between tools

RoomII

Palantir

The demo behind an NVIDIA × Lowe's partnership.

Highlights

  • One of 22 Meritocracy Fellows selected from 800 applicants.
  • Built the full Palantir × NVIDIA demo for Lowe's, which helped secure a multi-million-dollar partnership.
  • Built an automatic property-quoting workflow for insurance clients that made quoting 250% more efficient.
Meritocracy Fellows
22 / 800
quoting efficiency
250%
partnership supported
$M+
Palantir and NVIDIA Optimize Lowe's OperationsFilm

RoomIII

blnd

Your taste. Their taste. One perfect pick.

Highlights

  • Swipe-to-rate onboarding builds a taste profile in minutes.
  • Group Blends merge several profiles into one shared ranking.
  • A personal For You feed, informed by friends' taste.
  • One-tap Letterboxd import.
  • blnd app screen showing a group blend of movie picks
    Fig. 1Blends
  • blnd app screen showing a user's taste profile
    Fig. 2Profile
  • blnd app screen with recommendations
    Fig. 3For You

RoomIV

PepprAI (YC W25)

A meeting agent that actually remembers.

Highlights

  • Built a RAG meeting agent on Recall AI, Deepgram and Pydantic AI.
  • Built ingestion pipelines for Slack, Google Drive and Discord.

RoomV

Drawings

Perspective drawing taught me to look at structure before surface.

  • Plate 1Town by the SeaDigital
  • Plate 2Two-Point StudyGraphite on paper
  • Plate 3House Under the RidgeGraphite on paper
  • Plate 4Interior StudyGraphite on paper

RoomVI

Provenance

Where this work has been.

I was born in Queens, New York. When I was five, my family moved to Abu Dhabi. I grew up there speaking English and Mandarin, and picked up French and Arabic along the way.

For high school I went to The Lawrenceville School, a boarding school in New Jersey. There I ran the Big Data & AI Research Club, first-authored two papers on environmental machine learning, and was the starting kicker on varsity football.

After graduating I took a gap year and spent it shipping. I built a meeting agent at PepprAI (YC W25), was a Forward Deployed Engineer at Palantir, and clustered DGX Sparks at NVIDIA.

Now I'm at Northwestern studying Electrical Engineering and doing robotics research. I still keep a sketchbook. Perspective drawing taught me to look at structure before surface, and I bring that habit to the systems I build.

Journey

  1. Born

    Queens

    New York City

    40.73° N, 73.79° W

    Where it started.

  2. Age 5

    Abu Dhabi

    United Arab Emirates

    24.45° N, 54.38° E

    Grew up here, between four languages.

  3. 2021 – 2025

    Lawrenceville

    New Jersey

    40.30° N, 74.73° W

    Boarding school, research, football.

  4. 2025 – 2026

    Gap year

    Cupertino · New York · Santa Clara

    37.32° N, 122.03° W

    PepprAI, then Palantir, then NVIDIA.

  5. 2026 –

    Evanston

    Illinois

    42.05° N, 87.69° W

    Electrical Engineering and robotics at Northwestern.

RoomVII

Archive

Experience, publications and skills.

Experience

  1. 2026 – 2030B.S. Electrical EngineeringNorthwestern University
  2. 2026Software Engineer InternNVIDIA
  3. 2025 – 2026Forward Deployed Software EngineerPalantir
  4. 2025AI Researcher & Full-Stack EngineerPepprAI (YC W25)
  5. 2024 – 2025Founder & CEOSolve-AI.org
  6. 2021 – 2025High School DiplomaThe Lawrenceville School

Publications

  1. Evaluating the Relationship Between Socioeconomic Disparity and Potential PFAS Exposure in the United States with Machine Learning

    First author

  2. Predictive Modeling of Glyphosate Contamination Utilizing Random Forest Classifiers

    First author

Skills

Languages
  • Python
  • TypeScript
  • C/C++
  • Swift
  • Java
  • SQL
  • R
Systems
  • vLLM
  • CUDA graphs
  • RDMA / RoCE
  • Docker
  • Postgres · PGVector
Product
  • Next.js
  • React
  • FastAPI
  • Django
  • Pydantic AI
ML
  • PyTorch
  • scikit-learn
  • TensorFlow
  • RAG
  • Speculative decoding