Prithvi Sairaj Krishnan Seeking · summer 2027 · machine learning / software engineering

United States PatentUS 12,190,518 B1

[72]

Prithvi Sairaj Krishnan

Inventor · Austin, Texas

[54]Image-processing pneumonia detectionTitle of invention

[45]Granted January 2025Date of patent · age 18

[21]B.S. Electrical & Computer EngineeringUT Austin · December 2028

[56]3 papers · IEEE ×2, Harvard JEIReferences cited

[57]

A convolutional network reads a chest radiograph and reports whether it shows pneumonia, at 99% accuracy. I trained it in PyTorch, filed on the architecture, and presented the method to IEEE. I now break medical-imaging models rather than only build them: in Dr. John Virostko’s group at Dell Medical School I run robustness testing on Merlin, Stanford’s 3D CT foundation model. Along the way I taught 225 children to program and started three organisations. I am looking for a summer 2027 internship in machine learning or software engineering.

Prithvi Sairaj Krishnan — Fig. 1 — inventor
ECE, UT Austin
Class of 2028

prithvi.krishnan@utexas.edu · 512-796-8547 Full patent · Résumé · GitHub · LinkedIn · X

How a model of mine becomes something you can check

Fig. 02 · from training run to public record

01

Train

A convolutional network learns to read chest radiographs in PyTorch. Accuracy is measured on data it has never seen.

99% held-out

02

Write it up

The method becomes a paper, with the training procedure spelled out well enough for someone else to repeat it.

Full method stated

03

Defend it

Submitted to IEEE, accepted, and presented in person at ISNCC 2024 with the room asking questions.

IEEE 10759009

04

File it

An examiner at the USPTO reviews the architecture against prior art and either grants or refuses.

US 12,190,518 B1

05

Ship it

Where the work has a user, it goes online. The cost model runs as a live app anyone can try.

Deployed · public

Two of these steps needed someone else to agree.Square nodes = outside review

The record

Four pieces of work with a public identifier attached

Every entry below links to its primary source. Nothing here has to be taken on trust.

US 12,190,518 B1Sole inventor
May 2023 — Jan 2025
PyTorch · CNN

Image-processing pneumonia detection

United States Patent & Trademark Office · granted 2025

A convolutional network that reads a chest radiograph and reports whether it shows pneumonia. I trained it in PyTorch, filed on the architecture, and the patent was granted while I was eighteen. The work was written up for IEEE Xplore, shown at science fair, and picked up by three inventor podcasts including Inventors Helping Inventors.

99%detection accuracy

18age at grant

A second, smaller model built with students from Berkeley, Princeton and Harvard during the Inspirit AI internship reached 95%.

IEEE 10759009First author
2024
Cited on Xplore

Pneumonia detection, presented at ISNCC

IEEE Symposium on Networks, Computers & Communications

I presented the detector and its training method at ISNCC 2024 and answered for it in front of the room. The paper is indexed on IEEE Xplore and has been cited since.

IEEE 10924990First author
2024
C++ · simulation

Long-term effects of autonomous vehicles on transportation

IEEE Autonomous & Trusted Vehicles Conference

A modelling study, written in C++, of what happens to urban planning, traffic patterns and transport access as autonomous vehicles reach scale. It asks which of the promised effects survive contact with the infrastructure we already have.

JEI 24-251First author
2024
Python · live app

Predicting the cost of a medical procedure

Harvard’s Journal of Emerging Investigators · deployed

A regression model that estimates what a medical procedure will cost a patient, wrapped in a web app so the estimate is something a person can actually use. Built over a one-to-one research internship with Vivek Shankar, a senior ML engineer at YouTube, and taken through the full development lifecycle before publication.

Run itlive deployment

Current research

Finding where Stanford’s CT foundation model breaks

Merlin reads 3D abdominal CT and classifies across 1,692 conditions. My job in the lab is not to build it — it is to find the inputs that make it wrong.

Undergraduate Research Assistant · March 2026 — present
Dr. John Virostko’s group · Dell Medical School, UT Austin

Merlin is a vision-language model: it takes a CT volume and a clinical text report, and its published results cover phenotype classification, five-year risk prediction and report generation.

I built an edge-case suite across six categories, scored on cosine similarity, logit drift, KL divergence, top-3 rank retention, entropy and confidence gap. The first thing it turned up was not a degradation curve. It was a flat line.

Fifteen text variants against one fixed CT volume — correct, wrong, empty, negated, self-contradictory, a stock-market discussion, Mandarin, typos. The image–text similarity score moves exactly as you would hope. The predictions do not move at all.

Text impact · 15 variants, image held fixedabdominal CT

Text input	Sim	Drift	KL	Rank
Correct	0.3882	0.0000	0.0000	3/3
Wrong diagnosis	−0.0705	0.0000	0.0000	3/3
Empty string	0.2856	0.0000	0.0000	3/3
Negation	−0.2256	0.0000	0.0000	3/3
Adversarial	−0.0539	0.0000	0.0000	3/3
Finance text	0.0985	0.0000	0.0000	3/3
Mandarin	0.1294	0.0000	0.0000	3/3

Drift = mean absolute change in prediction logits against baseline. KL = shift in the full prediction distribution. Rank = top-3 conditions still matching baseline. Zero across every variant.

Finding 01

The language half of the vision-language model does not reach the prediction

Similarity swings from 0.39 to −0.07 as the text goes from correct to wrong, so the text encoder is working. But logit drift, KL divergence and top-3 rank do not move by any amount at all. For phenotype classification, Merlin is image-only at inference — and a clinician handing it a report would have no way to know that.

Finding 02

Noise robustness has a cliff, not a slope

A Rician-noise sweep across 15 sigma levels with 10 Monte Carlo seeds each — 150 inferences. Top-1 accuracy holds perfectly to σ≈0.05, then falls off entirely by σ≈0.10. Top-5 survives to σ≈0.15. The model does not degrade gracefully; it is right until it is abruptly not.

Tools

Instruments for making a system prove what it is doing

Most of what I have built this year asks one question in different forms: not whether a model performs, but whether anyone outside it can check. It is the same question as the Merlin work.

quantprint

Works out what an LLM API is really serving. It reads the top-k logprobs any endpoint already returns and identifies the model and the quantization behind them. A provider can swap fp16 for int4, change a kernel, or route to a smaller model without telling anyone — and today nobody outside the provider can prove it happened.

PythonSept 2026
GBDT · logprob features

StatusPipeline verified on synthetic traces. No real-model numbers yet.

loadbearing

Finds which parts of a system prompt are actually doing work. Evals tell you the score moved; they do not tell you which of the forty lines moved it. This runs a leave-one-out ablation over prompt segments and reports what each one costs you when deleted. It ships with a test that plants a known dependency and checks the tool recovers exactly that set — so the harness cannot quietly produce a confident ranking of noise.

PythonAug 2026
Ablation harness

StatusHarness validated, tests pass. No model sweep run yet.

Tally

An agent that files your expense report. It matches every card charge to the right receipt in Gmail, pulls the business purpose from Calendar, and escalates what it cannot resolve rather than guessing. The deterministic stages alone clear every target in the spec — which means the model has to beat that table before it earns a place in the pipeline.

0.906match accuracy · target 0.85

1.000escalation recall

0.000false confidence

PythonSept 2026
Gmail · Calendar · agent

StatusBaseline lifted 0.42 → 0.91. No model call in any figure shown.

signet

An e-signature API whose output proves itself. A completed PDF carries its original, every signature image, and a hash-chained audit record sealed with an Ed25519 signature. Anyone can verify it offline against the public key — forty lines, two dependencies, no account and no call back to the vendor. DocuSign sells trust in DocuSign; this sells a document that does not need any.

PythonSept 2026
FastAPI · Ed25519

StatusRunning. Offline verifier shipped.

primer

A reading tutor for ages four to six that can only say words the child has already been taught. Most AI reading apps generate a story and hope. This one checks every single word against a hand-verified phonics bank and rewrites until it passes — reject dinosaur, reject mother, try again. The constraint is the product, and it comes straight out of two years of teaching children to read code.

JavaScriptAug 2026
Constrained generation

StatusRunning.

Also on GitHub — igdm, shadow-mode Instagram automation that diffs itself against the incumbent before it sends anything · GridOS, energy interconnection tooling · SnapCycle · Echo · Insurance-Premium-Predictor, the model behind the JEI paper

Experience

Two things at once

From 2023 to 2025 the left column and the right column were happening in the same weeks.

Research and employment

UT Austin · Dell Medical School

Undergraduate Research Assistant · Mar 2026 — present

Robustness testing of Merlin, Stanford’s 3D CT foundation model, in Dr. John Virostko’s group. Edge-case suite across six categories; found the model’s text input has no effect on its phenotype predictions.

Deep24 Builder Fellowship

Builder Fellow · May — Aug 2026

Built LedgrAI, an AI personal-finance coach for first-generation earners. Plaid bank connection, a 0–100 “Money Pulse” read on 30 days of transactions, a Sunday brief, and a dollar spending plan generated when a paycheck lands. Solo build in React Native, PRD to App Store submission in four weeks.

The Global Career Accelerator

Coding for Data Intern · Jan 2026 — present

Data and product analytics run against real records from Intel, OpenAI and the Recording Academy, in SQL, Python, Tableau and R.

Hey, Blue!

Web App Engineering Intern · Jul 2024 — Aug 2025

Angular product features end to end, backend infrastructure and API endpoints, authentication reworked. Delivered against requirements set with Walmart and Panasonic.

Mathnasium

Mathematics Instructor · Jan 2024 — Jun 2025

200 students across ten grade levels, and the progress notes their parents read each week.

Founded and ran

Trail2Tech

Founder & director · Jan 2023 — Aug 2025

Coding and AI for grades one to six. I wrote the curriculum and took the programme to campuses outside Austin.

Code Club Austin

Founder & director · Jan 2023 — Aug 2025

The Austin campus. Python and Java, with the parents kept in the loop. KUT covered it on air.

Dev0ne Design

Founder · Jan 2023 — Aug 2025

Custom websites for small businesses around Austin, from the first conversation through design, build and deployment.

225 children were taught in the same period I was interning, instructing, and publishing three papers.

Teaching and community

Two nonprofits and a studio

Founded and run 2023 to 2025, alongside the research.

I wrote the curriculum

AI, Python and Java for grades one through six — sequenced so a seven-year-old and an eleven-year-old could sit in the same room. 225 children went through it.

I ran the operations

Enrolment, scheduling, instructors and the weekly note home to parents, across two campuses.

It got covered

KUT Austin put the Code Club campus on air.

And a studio alongside it

Dev0ne Design built websites for small businesses around Austin, first conversation through deployment.

Programme record2023 — 2025

students_taught225

grade_range1 – 6

organisations3

campusesmultiple

curriculumAI · Python · Java

pressKUT Austin

run_alongside3 papers, 1 patent

Honours

Three that carry weight, and four more

Ordered by what they took, not by date.

2023

USACO Gold Division

Promoted to Gold in the USA Computing Olympiad, out of roughly 4,000 competitors.

2023

Apple Swift Student Challenge

Winner. Apple selects worldwide on the strength of a playground built in three weeks.

2025

IEEE Programme Committee

Invited to review submissions for IEEE Xplore conferences, including CICN 2025 — at nineteen.

AIME qualifier2023 · through the AMC

Stanford International Logic Olympiad2023 · semi-finalist

President’s Volunteer Service Award, Gold2022–25 · received twice

Jane Street probability puzzle2025 · exact Markov chain

Service

Eagle Scout service project

Trail signage for a Cedar Park nature preserve, July 2022 to February 2023. Eleven years of Scouting to get there, and eleven of karate to black belt alongside it.

01

Scope

Surveyed five miles of trail and worked out what signage the preserve actually needed.

5 miles surveyed

02

Approval

Took the plan to Cedar Park Parks & Recreation and revised it until it met city guidelines.

City sign-off required

03

Funding

Pitched Lowe’s and secured materials funding for the entire build.

Fully funded

04

Build

Led 20 volunteers through construction and installation across the preserve.

20+ volunteers · 250+ hours

Ask

Rather ask than read?

The patent, all three papers and the résumé are loaded. Pick a question or write your own — it cites whichever source it answered from, and says so when it does not know.

Ask Prithvi’s work4 sources indexed

Answer

A convolutional network that reads a chest radiograph and reports whether it shows pneumonia, at 99% accuracy on held-out data. The patent covers the architecture. It was trained in PyTorch and filed while Prithvi was seventeen; the grant came through in January 2025, at eighteen.

The same work was written up for IEEE and presented at ISNCC 2024.

Sources
US 12,190,518 B1 · USPTO
IEEE 10759009 · ISNCC 2024

PythonJavaC++CTypeScriptJavaScriptSQLRC#RustARM AssemblyLC-3PyTorchReactNext.jsAngularFastAPIAWSTableauGit

I am looking for a summer 2027 internship.

Machine learning, applied research, or software engineering. Austin or anywhere. If any of the four things above were interesting, the fastest way to find out more is to email me.

prithvi.krishnan@utexas.edu Résumé GitHub LinkedIn X

B.S. Electrical & Computer Engineering · UT Austin · December 2028 Austin, Texas · 512-796-8547 · English, Tamil, Spanish, Mandarin