Pranjal
Mishra

I work on machine learning and agentic AI in healthcare.

Research Assistant at NYU Langone Radiology. First author on an ISMRM abstract presented in Boston. MS in Computer Science at NYU, graduating May 2027.

NowNYU Langone Radiology

RadAssistant

An agentic assistant inside 3D Slicer. A plain-English request becomes a plan that runs across 50 validated imaging tools.

Every number it reports traces back to a tool result, a deterministic verifier checks each step, and consequential actions wait for a person. The lab's radiologist uses it day to day.

LangGraphTool callingEvalsHITL3D Slicer
3rd of 4,000+ teamsAMD Pervasive AI Contest

AudioVision

Generative restoration for historical footage: colorization, 4x super-resolution and generated sound, running as GPU inference on AMD hardware.

I led a team of five and we demoed it live on AMD's keynote stage.

PyTorchROCmGAN colorizationESRGANAudioCraft
2nd of 5,000+ teamsShell.ai Hackathon

Fleet-Opt

The product around a CPLEX optimizer that plans a fleet's move from diesel to electric over a 16-year horizon.

Next.js front end, Express and Prisma on PostgreSQL, and a Redis job queue with batching and pooling, so long solves stream progress instead of blocking the interface.

Next.jsExpressPrismaRedisCPLEX
2026Build with AI Hackathon · Google x NYU

VoiceGraph

A voice agent over a Neo4j knowledge graph. Ask out loud and it picks from 8 retrieval strategies, from text-to-Cypher to community detection.

Built solo overnight: Gemini Live, FastAPI, and a React and Three.js graph that redraws as it answers.

Gemini LiveNeo4jFastAPIThree.jsCloud Run
2026Live on AWS

Research Paper Summarizer

A serverless AWS pipeline that summarizes an arXiv paper in about 30 seconds, with cited chat and a knowledge graph per paper.

Step Functions map-reduce over Lambda, content-hash dedup in DynamoDB, and a quota that reserves before it spends. It costs nothing while idle.

AWS CDKStep FunctionsBedrockDynamoDBNext.js
Work

Also

Open source

Eight merged pull requests

kornia, the PyTorch computer vision library, and career-ops. Each one started by reproducing the bug. Each fix ships with a test that fails without it.

3kornia
5career-ops
1,011lines merged
All merged PRs on GitHub
  1. kornia #466911k stars

    Drop quartic roots that do not satisfy the quartic

    Three review rounds. Property tests against numpy as ground truth, then mutation testing so every tolerance has a test that kills it.

    Sep 24
  2. kornia #467511k stars

    Rasterise all RandomRain drops in one indexed write

    8.8x faster at batch 32 on Apple silicon, bit-exact to the old output across 12,482 configurations.

    Sep 20
  3. kornia #466811k stars

    Reject non-finite explicit ranges in _range_bound

    An infinite endpoint slipped through only when the default domain was itself infinite. Found by reading the comparison, not the issue.

    Sep 20
  4. career-ops #430672k stars

    Recognise section headings in every language the repo ships

    Non-English resumes were failing the ATS gate outright. Fixed 18 of 19 languages with the repo's own vocabulary tables.

    Sep 25
  5. career-ops #443272k stars

    Find the draft block by a trailing (draft) marker

    One English heading was the only match, so 16 languages silently regenerated answers they had already paid for.

    Sep 25
  6. career-ops #427072k stars

    Stop the shipped template's own CSS from docking 15 points

    Three false positives in the ATS scorer. Rendered resume score went from 84 to 99 of 100.

    Sep 18
  7. career-ops #428372k stars

    Trim CSS whitespace so U+00A0 cannot hide a font name

    The follow-up the maintainer handed over after the first merge.

    Sep 18
  8. career-ops #430572k stars

    Brace the escape for a character above the BMP

    Residue from the same review thread, taken and landed the same week.

    Sep 23
Presented at ISMRM 2026

Research

PresentedISMRM 2026 · Boston

Multimodal MRI biomarkers for muscle aging

First and presenting author, Workshop on Musculoskeletal MRI

I built the comparative ML pipeline in Python and R: 11 model families benchmarked over 6,752 radiomics features per leg, with subject-level cross-validation, 5,000-sample cluster bootstrap and permutation testing on NYU's HPC cluster. Texture features lifted strength-prediction R² by 0.17 to 0.21 over volume baselines. A JMRI manuscript is in progress.

Abstract
PaperNYU · 2025

Scaling laws for music language models

Course research

Trained 9 transformer and LSTM models from 1.7M to 92.9M parameters on 178,561 MIDI files. Music scales steeper than language (alpha 0.19 against 0.076), and transformers were 2.4x more parameter-efficient than LSTMs at 90M.

Code and paper

Experience

Newest first.

Résumé
2026
NYU Langone Health, RadiologyResearch Assistant

RadAssistant, a first-author ISMRM abstract presented in Boston, and the clinical app for an NCI-funded prostate cancer AI project.

2025
NYU TandonMS Computer Science, 4.0

Machine learning, deep learning, generative AI, cloud computing, algorithms. LeetCode Bootcamp badge

2025
CASDIC, DRDOML Engineering Intern

Radar signature classifiers, up 15% in detection accuracy, and a reinforcement learning agent, up 18% on a manual protocol. Mostly under NDA.

2024
Two competitionsAMD Pervasive AI, Shell.ai

3rd of 4,000+ teams at AMD. 2nd of 5,000+ at Shell.ai.

2023
PRL, ISROSoftware Engineering Intern

A Kafka pipeline across 21 LIDAR stations, hundreds of GB a day, replacing hard drives moved by hand. My first LIDAR.

Contact

Get in touch

Email is the fastest way to reach me. I answer every serious message. Or pick a slot and we talk.

Book a call30 min · Google Meet