Harsh Patel
Developer + Student, Builder, Tinkerer, and ML Engineer
Semantic caching for LLM APIs
A FastAPI gateway that answers repeat questions from a pgvector cache instead of paying for another LLM call. Cache hits come back about 100× faster, and I measured how often they answer the wrong question.
Hand-written CUDA kernels for attention
Five kernels, from a vector add up to a fused flash-attention-style kernel, each benchmarked against PyTorch and cuBLAS on a T4. The write-up covers what I got wrong along the way.
What the kernels run on
A look inside a graphics card: the chip, its memory, the power stages, and the cooler that sits on top. Keep scrolling to lift the lid.
Card binders, rebuilt for iOS
BinderDex is an iOS app where collectors digitally recreate their physical card binders, with smooth search across 19,000+ cards. Built with SwiftUI, SwiftData, and StoreKit 2 subscriptions.
Serving an LLM to a crowd
An inference server I’m building in stages, from a one-request-at-a-time baseline toward continuous batching. Every stage is benchmarked on time to first token and throughput.
By the numbers
- 102×
- faster response on a cache hitLLM Gateway
- 1.60×
- speedup from shared-memory tilingCUDA Kernels
- 19,000+
- cards to search and collectBinderDex
- 14+
- feature domains built as sole engineerSELEST Security
Where I’ve worked
Software Engineering Intern
SELEST Security- Node.js
- Express
- Prisma
- PayPal
- Supabase
- Next.js
- Designed, built, and deployed SELEST’s payment and subscription system end-to-end, implementing PayPal recurring billing with webhook-driven lifecycle handling (renewals, cancellations, failed payments) on a Node.js/Express backend with Prisma ORM.
- Owned full-stack development as the sole engineer, architecting 14+ feature domains (authentication, payments, messaging, bookings, reviews) and rebuilding the mobile frontend from React Native/Expo to a Next.js + Capacitor architecture.
- Built secure account access and authentication using Supabase, covering login, session management, and account security across web and mobile.
Software Engineering Intern
GreenVeil Packaging- Shopify
- Liquid
- JavaScript
- Responsive design
- Built asynchronous filtering and sorting for a fast-growing cleantech company’s production Shopify storefront, using JavaScript and Liquid to fetch and re-render product results without full-page reloads, and shipped new price and category filters.
- Refactored the product listing page and grid into reusable Liquid sections and snippets, restructuring layout and visual hierarchy to improve product discoverability.
- Rebuilt responsive layouts across mobile, tablet, and desktop breakpoints, resolving layout defects and delivering a consistent shopping experience on every device.
Software Engineering Intern
Appy.yo Innovation Lab- Authentication
- Databases
- UI components
- Railway
- Built production features for early-stage web apps: authentication flows, database integration, and UI components.
- Tested, debugged, and prepared deployments, supporting live launches of apps used by small and medium-sized businesses.
- Deployed and launched applications on Railway.
Graphic Design and Junior Software Engineering Intern (Co-op)
Signarama Brampton- Graphic design
- Workflow automation
- Frontend
- Designed and produced digital and print graphics for signage, branding, and marketing materials.
- Helped with internal tools, workflow automation, and basic frontend web development.
- Streamlined file preparation to reduce errors and speed up production.
Western University
London, Ontario- BSc, Honours Specialization in Computer Science
- Ivey Business AEO
- Python
- Java
- C
- C++
- JavaScript
- SQL
- CUDA C++
- React
- Next.js
- Node.js/Express
- React Native
- PyTorch
- Pandas
- NumPy
- FastAPI
- Git
- Prisma
- Supabase
- PostgreSQL
- Google Colab
- Jupyter
- Shopify Liquid
- Railway
- Docker
- Redis
Hi there
I’m Harsh, a computer science student at Western University in London, Ontario. The newest challenge I’ve been trying to tackle is making LLMs fast to serve, from caching and batching requests down to writing the GPU kernels underneath.
Away from the keyboard I love sports and music, and both keep finding their way into my code. The sports side became an NBA player stat projection model that pulls historical stats and projects performance across seasons and matchups with pandas. The music side became a Spotify Wrapped clone that uses the Spotify API to show top artists, tracks, genres, and listening time.
I’m always up for hearing about new projects and opportunities. You can reach me at harshptl0715@gmail.com or 437-980-1956.
Get in touch