Ishaani
Jain
CS & Physics @ UC Irvine, building at the intersection of systems programming, ML infrastructure, and research.
Builder. Researcher.
Leader.
Hi, I'm Ishaani, a freshman at UC Irvine with junior standing, studying Computer Science and Physics. I care deeply about building technology that makes a real difference and love working on hard problems at the systems and ML layer. Currently doing CV research at Calit2, ML infrastructure at GoFlyy, and biomedical data tools at Oregon State's Ramsey Lab.
Where I study
University of California, Irvine
B.S. Computer Science & Physics · Expected Jun 2027
– Jun 2027
Where I've worked
- Building high-performance Python/C++ tools for NIH NCATS to ingest & query the Babel concept identifier normalization database for the Biomedical Data Translator program
- Developed optimized scripts for large-scale biomedical data translation
- Built physics-based CV algorithm: 100% accuracy, 0 false alarms on sub-pixel fires at 6.8–10.7 mi; deployed for ALERTCalifornia's 1,000+ camera network
- Engineered hyperspectral pipeline: 500-band datacubes (6,788×1,240 px / 2 min) via 1.4 Gbps link with multi-modal fusion (VNIR / SWIR / MWIR / LWIR)
- Classified 13+ species (CH₄, CO₂, NO₂, PM2.5) across 40 km for real-time air quality & fire monitoring
- Building CV pipeline for automated garment condition assessment with 90%+ defect classification accuracy
- Optimized inference latency 67% (2.1 s → 680 ms) via TensorRT INT8 quantization
- Developing virtual try-on system to reduce return rates
- Architected ML extraction pipeline: 90%+ accuracy on aviation compliance docs via AWS Bedrock; Lambda cold-start optimized 65% (8 s → 2.8 s)
- Serving 2 international clients across 500+ documents processed, 90%+ test coverage across 15K LOC
- 8-week fellowship on reward misspecification, deceptive alignment, and instrumental convergence in vision-language models
- Red-teamed GPT-4, Claude 3.5, and Gemini Pro for deceptive capabilities; investigated reward hacking & scalable oversight techniques
- Led 3-engineer team to win 1st place ($10K seed) out of 11 teams
- Architected full-stack Next.js / Firebase startup matchmaking platform with AI semantic search
Selected work
Candid: Civic Tech RAG Platform
Tells you exactly how ballot measures affect your wallet, personalized to your ZIP, income, and housing situation. 52,500+ real legislative chunks in ChromaDB; linear regression on 2,151 rows of Census Bureau data (R² = 0.61); live 3D Mapbox map; real politician voting records from OpenStates; Groq LLaMA 3.3 70B reranking against your personal profile.
CUDA GEMM Kernels
Matrix multiply from scratch across 4 optimization levels: naive, shared-memory tiling, register blocking, and warp primitives. Reaches 86% of cuBLAS at large sizes. Profiled with Nsight Compute; exposed as a PyTorch C++ extension.
Document Genie
ML extraction pipeline for aviation compliance docs with 90%+ accuracy, Lambda cold-start cut 65% (8 s to 2.8 s). Serving 2 international clients processing 500+ documents with 90%+ test coverage across 15K LOC.
Calendar AI
Turns plain English into Google Calendar events. "Study for 2 hours" → optimal slot with no conflicts, auto-synced. Reads real calendar context so there's never a double-booking.
CNN Image Classifier
ResNet-18 on CIFAR-10 (60K images, 10 classes) with data augmentation, LR scheduling, and early stopping. 90% test accuracy. Deployed with ONNX for inference.
Startup Matchmaker
Matchmaking platform for startups and investors with AI semantic search. Led 3-engineer team to win 1st place out of 11 teams at C2S Technologies' internal pitch competition.