Hey!!! I'm Ishaani, a Computer Science student at UC Irvine who cares deeply about building technology that makes a real difference. Here's a bit more about me:
💖 Currently I'm doing systems optimization research at MIT CSAIL, generative AI research at the CMU Human Sensing Lab, and SWE at C2S Technologies.
🌸 My focus is bridging the gap between low-level performance optimization (CUDA, C++) and high-level AI (LLMs, PyTorch) to create robust, impactful solutions!
💖 Hackathons: Recently won 1st place at IrvineHacks with Candid, an app that tells you exactly how ballot measures affect your wallet using RAG pipelines and ChromaDB.
🌸 Interests: I love working on hard problems at the systems and ML layer, especially parallel computing, computer vision, and AI safety!
💖 Reach me via my LinkedIn or Email! 🌸 (Bonus: if you can't already tell, My favorite color RN is purple, it changes)
Programming & Core Concepts
Machine Learning & Data
Backend & Cloud
Massachusetts Institute of Technology (CSAIL) — Researcher
- Researching at the intersection of high-performance computing (HPC) and computer graphics to develop and optimize GPUs.
- Accelerating irregular and unstructured computational problems by designing efficient GPU systems and programming models in C++ and CUDA.
- Optimizing memory hierarchies and execution pipelines for high-performance computing applications.
Carnegie Mellon University (Human Sensing Lab) — Researcher
- Working under Professor Fernando De la Torre on research at the intersection of computer vision, human sensing, and generative AI.
- Authoring a photorealistic Unreal Engine 5 synthetic data pipeline generating large-scale annotated human motion sequences.
- Designing multi-scenario video generation protocols to quantify how synthetic visual diversity drives pose estimation accuracy in real-world robot deployments.
Oregon State University (Ramsey Lab) — Undergrad Researcher
- Architecting high-performance infrastructure for the NIH NCATS Biomedical Data Translator, directly improving how scientists query and integrate global medical databases (Gov funded).
UC Irvine (Calit2 & SPAR) — Undergrad AI/ML Researcher & AI Safety Fellow
- Built physics-based AI to detect wildfires from over 6 miles away, providing critical early-warning data to ALERTCalifornia.
- Selected for a competitive fellowship to conduct specialized research on AI evaluation methodologies and alignment.
Software Developer Engineer Intern @ C2S Technologies, Inc.
- Developing full-stack software solutions and maintaining enterprise-grade applications focusing on usability and performance.
- Previously led a team of three to win 1st place and $10,000 in funding for an AI-driven investment matchmaking platform using Retrieval-Augmented Generation (RAG) and UED.
Software Development Intern @ GoFlyy
- Developed CV-driven automation and routing logic that transformed manual clothing inspections into an efficient, AI-powered logistics pipeline.
- Achieved 90%+ defect classification accuracy and slashed inference latency 67% (2.1s to 680ms) via TensorRT INT8 quantization.
Founding Software Engineering Intern @ Stealth Aviation Tech Startup
- Scaled an AI document pipeline to 94% accuracy and sub-3-second processing, directly securing two international aviation clients for pilot deployment.
- Built production full-stack infrastructure on AWS (Lambda, DynamoDB, S3, CDK) to process 500+ documents/month.
Outreach Officer @ AWSCloud@UCI
- Connecting students with cloud computing resources, organizing workshops, and facilitating AWS certification study groups.
| Project | Tech Stack | Description |
|---|---|---|
| Candid (IrvineHacks Winner) |
   |
Tells you how state and federal ballot measures affect your finances with a real, calculated number, personalized to your ZIP code and income. Processed 52,500+ legislative chunks into ChromaDB vector storage. |
| CUDA Kernels for Matrix Mul |   |
Built from scratch with four optimization levels (naive, shared-memory tiled, register-blocked, warp-optimized). Reaches 86% of cuBLAS. Profiled with Nsight Compute and integrated as a PyTorch C++ extension. |
| ResNet-18 Image Classification | |
Trained a ResNet-18 on CIFAR-10, achieving ~90% accuracy. Modified the architecture for 32x32 images by swapping the first conv layer for a smaller 3x3 one and removing the initial maxpool to prevent aggressive downsampling. |
| DataLoader Optimization |  |
Created a hands-on toolkit for profiling and optimizing PyTorch data loading pipelines. Demonstrated how to cut data loading overhead from ~40% to ~10% of total training time using smart batching and CUDA prefetching. |
| Document Genie & Calendar AI |   |
Calendar AI: Turns plain English into Google Calendar events with smart collision detection. Document Genie: ML extraction pipeline for aviation docs with 94%+ accuracy; slashed Lambda cold-start latency by 65%. |