Hi, I’m Dipesh.

eevsec.com · school cyber-safety & computational-thinking platform, Classes 3-10

• Sole engineer on Cyber Suraksha, a cyber-safety and computational-thinking platform for Classes 3-10 now onboarding 3 schools; built on Next.js, TypeScript, and Cloudflare Workers over a 19-table multi-tenant Cloudflare D1 schema, with PBKDF2 authentication, hashed server-side sessions, and role-scoped student, teacher, and school-admin dashboards.
• Shipped a Class 10 mathematics course spanning 18 interactive topics, 9 narrated animated lessons, and 95 auto-graded questions, plus 4 gamified Class 3 modules; grading runs server-side against a version-controlled question bank, with a single-attempt rule enforced by a database uniqueness constraint so marks cannot be replayed or forged from the client.
• Engineered a retrieval-grounded AI assistant that answers student questions from course handbooks with inline source citations, plus progress rollups that report per-topic and per-chapter mastery to teachers in a single indexed read instead of scanning raw attempt history.
• Hardened delivery with 45 automated tests and browser gates covering auth boundaries, grading integrity, layout regressions, and cross-browser media playback on Chromium and WebKit.
• Designed and shipped eevsec.com end to end, a 20-page product marketing site covering product, pricing, and audience pages with responsive dark-mode layouts and search-optimized metadata, driving 500+ waitlist signups before launch.

• Next.js
• TypeScript
• Cloudflare Workers
• Cloudflare D1
• Multi-Tenant Auth
• RAG
• Playwright
• EdTech

New York, USA · two-photon neuron segmentation & cross-imaging matching

• Built a two-photon neuron-segmentation pipeline for 3D reconstruction, converting 7,407 VIA polygons into Cellpose masks and linking z-slice annotations into 2,548 tracked cells for registration and model training.
• Developed a soma-print spatial-registration model with Fused Gromov-Wasserstein optimal transport to match neurons across in-vivo and ex-vivo imaging, replacing ICP; achieved 95% rank-1 matching without rigid pre-alignment and 92% pooled rank-1 across datasets of 873 and 691 cells.
• Fine-tuned Cellpose-SAM on 7,407 annotations from 75 two-photon images, raising leakage-controlled held-out ex-vivo AP@50 from 0.14 to 0.55.
• Trained a unified Cellpose-SAM across 7 datasets with distributed GPUs, raising cross-dataset AP@50 from 0.28 to 0.40; reduced out-of-domain failure and standardized annotations with neuroscience researchers.

• Cellpose-SAM
• PyTorch
• Optimal Transport
• Image Registration
• Instance Segmentation
• Distributed GPU
• HPC

Halifax, Canada · resume-job relevance modeling

• Built an HR resume screener that assigns resume-job pairs to 3 relevance classes; trained a PyTorch CNN-Transformer with DistilBERT embeddings and 1D convolutions, reaching 94% held-out validation accuracy.
• Generated weak labels by pairing 1,962 resumes with 1.6M+ job postings using TF-IDF skill similarity; used synonym augmentation, resampling, and loss tuning to reduce false positives and prioritize high-fit candidates.

• DistilBERT
• PyTorch
• 1D CNN
• NLP
• Weak Supervision

Bengaluru, India · aerial object detection

• Developed and fine-tuned an Oriented R-CNN to identify and classify small, arbitrarily rotated ground targets in aerial imagery using Midpoint Offset Representation and rotated RoI Align.
• Optimized 1024×1024 inference with model pruning and CUDA parallelism, reaching 92.8% region-proposal recall and 15.1 FPS for near-real-time surveillance.

• Oriented R-CNN
• CUDA
• Object Detection
• Model Pruning
• OpenCV

Master of Science in Computer Science · GPA 3.89 / 4.0 · New York, USA

Coursework: Computer Vision · Machine Learning · Foundations of Data Science · Neuroinformatics

Bachelor of Technology in Computer Science and Engineering · GPA 8.41 / 10 · Chennai, India