PRS Prithish Shan

Prithish Shan

Building across biotech and AI to deliver solutions into the hands of end users.

Education
Carnegie Mellon University
B.S. Statistics & Machine Learning, Computer Science
Contact
pshan@andrew.cmu.edu
507-271-2799
Jacksonville, FL · Pittsburgh, PA
Links
github.com/prithishshan
linkedin.com/in/prithish-shan

Scroll to explore the work

MyoVerse

MyoVerse mobile app
MyoVerse EMG sensors
MyoVerse backend architecture
MyoVerse analytics website
MyoVerse landing page

Biomedical AI

Presenting RNA splicing research at Mayo Clinic
Dermal: acne lesion detection on iOS
Renal Stone AI: segmentation of Randall's plaque and stone deposits on the renal papilla Work in progress

Agentic Apps

GeoBlend: Congressional App Challenge at the U.S. Capitol
Itera: ProdHacks 2026, SAP Intelligence Track
PriorAuth AI: app UI and Azure architecture
Real-Estate AI: parcel map & compliance copilot

Foundational AI

Mosaic Grad: tensor matmul kernel in pure C
Mosaic GPU: distributed render across laptop and phone worker nodes
Foundational · 03

Mosaic GPT

Transformer from scratch

Work in progress
∇
Foundational · 04

Mosaic CAD

Programmatic geometry kernel

Work in progress
∇

MyoVerse

A wireless surface-electromyography (sEMG) platform for sports physical therapy. I designed, deployed, and continue to maintain the software infrastructure from the ground up, including the device firmware, the mobile app, the cloud API, and the analytics web surface.

01

Mobile App

Flutter app for seamless Bluetooth connection and assignment of sensors onto their respective muscles, with controller-maintained state for reactivity. Isolated thread workers run real-time digital signal processing while OAuth2 handles identity and every session streams to the cloud for data and long-term insights.

DartFlutterBLEDSP
02

Sensor Firmware

C device firmware on the Zephyr RTOS platform for robust wireless timekeeping and transmission of high-frequency EMG data packets.

CZephyr RTOSnRF24BLE GATT
03

Data API

A TypeScript API service deployed on AWS Lambda that stores 2 million-plus sensor logs in real time and serves long-term analytics to web and mobile consumers.

TypeScriptLambdaRDS PostgresSQSS3
04

Analytics Website

A Next.js data analytics platform where users export session data and monitor recovery through biomarker trends.

Next.jsTypeScript
05

Landing Page

The startup's public landing page, the front door for the sensing platform.

Next.jsTypeScript
Visit live ↗

Biomedical AI

Machine learning applied to real clinical problems at Mayo Clinic.

01

Van Blitterswijk Lab, Mayo Clinic

Built an ML pipeline characterizing novel RNA splicing in genetic ALS and FTLD from long-read mRNA sequencing. Automated HPC workflows with shell scripts and trained MLPs, SVMs, and random forests with Bayesian tuning and cross-validation to identify prognostic transcripts and predict clinical outcomes from unseen patient data. Presented a poster at Mayo Clinic, regional, and state conferences.

PythonTensorFlowscikit-learnMLP / SVM / RFBayesian tuningHPC
View poster ↗

Dermal: Acne Analysis AI

A production iOS app (Swift) with a FastAPI backend that segments acne lesions from user photos using a U-Net trained on 1,200 dermatologist-annotated images on Azure ML, then applies optical-density physics to map hemoglobin and melanin distribution for lesion-type classification and quantification of abnormal pigmentation.

SwiftFastAPIU-NetAzure MLPyTorch
03

Renal Stone AI

Segmentation of intrarenal calcification from endoscopic renal-papilla imagery, outlining Randall's plaque and stone deposits to quantify calcification burden.

PythonSegmentationMedical imaging
Work in progress

Foundational AI

The Mosaic suite explores the primitives of modern AI from the ground up: automatic differentiation, distributed GPU compute, language models, and geometry.

01

Mosaic Grad

A tensor + automatic-differentiation engine in pure C, no framework. Tensors carry shape, strides, and gradients; operations record onto a computation graph that a topological backward pass differentiates in reverse mode. Was able to train an MLP classifier on the MNIST OCR dataset to 97.6% accuracy.

CAutogradTensors
View on GitHub ↗
02

Mosaic GPU: Distributed GPU Computing

A response to expensive GPU compute: split a parallelizable GPU workload into independent jobs and farm them out to clients who contribute from a plain browser tab. A Next.js app, a Postgres database with concurrency handling for job claiming, and centralized artifact storage in S3.

Next.jsTypeScriptPostgresAWS S3
View on GitHub ↗
03

Mosaic GPT

A GPT-style transformer built on PyTorch, with the attention, tokenization, and training loop implemented directly on top of it rather than pulled from a high-level library. Work in progress.

PyTorchPythonTransformerAttention
Work in progress
04

Mosaic CAD

An early-stage CAD foundation model that treats parametric CAD generation as an autoregressive process, with CAD construction instructions (sketch, extrude, boolean) as the output space rather than pixels or Python. A B-rep encoder gives the model true context on the geometry it is building, and a Qwen-based encoder conditions on the prompt. The aim is CAD generation usable inside an agentic loop, without the context and precision that current solutions lose by feeding agents screenshots and taking back Python scripts.

PyTorchQwenAutoregressiveB-repPythonOCC
Work in progress

Agentic Apps

AI products shipped to real users and built under hackathon deadlines, from award-winning apps to agentic platforms.

01

GeoBlend

An app that teaches users about nearby historical sites via GPS, serving real-time streamed output from the OpenAI API. Built with React Native + Expo and a Node.js/TypeScript API. Won the Congressional App Challenge for Florida's 4th district and was presented at the U.S. Capitol.

React NativeExpoNode.jsOpenAI APIGPS
Congressional App Challenge ↗
02

Itera: SAP Intelligence Track, 1st Prize

A generative-AI ideation platform that turns natural language into editable 3D models: image → generative 3D (HuggingFace) → OBJ → Autodesk API → an editable BRep in an in-browser 3D viewer, so a non-technical PM can refine a concept by describing changes. Won the SAP Intelligence Track.

Next.jsGenerative AIHuggingFaceAutodesk API
1st Prize · SAP
03

PriorAuth AI

A prior-authorization agent that tells a clinician, at the moment an order is placed, whether it will clear the payer's policy and what documentation is missing, instead of days later in a denial letter. It compiles each insurer's medical policy into a versioned criteria tree where every node cites its source, reads the patient's record over FHIR (problems, meds, labs, vitals, notes), evaluates each criterion against the evidence, and returns a cited determination with the specific fixes needed. Built on Foundry IQ for HIPAA-compliant access to frontier models, AI Search, and Postgres/Blob, keeping PHI in-tenant and integrating over CDS Hooks and SMART on FHIR. Validated the integration end to end by registering the SMART on FHIR app in the Epic Sandbox and running the standard EHR launch and OAuth2 flow to pull live sandbox patient records into the agent.

Next.jsTypeScriptFoundry IQOpenAIFHIRSMART on FHIRCDS HooksEpic sandbox
04

Real-Estate AI: Agentic Feasibility Platform

An agentic real-estate feasibility platform using RAG for land discovery and analysis. A geospatial pipeline ingests 100k+ land parcels into PostGIS and indexes zoning documents as pgvector embeddings; a scalable Express.js API wires LangChain agents into an Angular frontend. Validated through customer discovery with commercial real-estate developers, including the CEO of RISE Commercial.

AngularExpress.jsLangChainPostGISpgvectorRAG