Building reliable backend systems, applied machine learning pipelines, and computer vision solutions that solve real problems.

I design and build production-oriented backend systems, low-latency ML pipelines, and computer vision solutions with a strong focus on reliability, scalability, and real-world constraints.

Experience

• Data Analytics & AI Intern — Tech Mahindra (Jun–Jul 2026)
  Worked on telecom domain analytics and AI-driven optimization. Built data pipelines and ML components for classification & anomaly detection on large-scale operational datasets under strict enterprise data governance.
• Research Intern — Samsung PRISM (AI/ML & 5G Network Systems)
  Engineered AI-driven 5G RAN scheduler in PyTorch achieving 71,000+ scheduling decisions/sec with sub-10 µs inference latency. Optimized C–Python shared-memory IPC and delivered 17× speedup via vectorization.
• Undergraduate Researcher — Medical AI, MIT Manipal
  Hybrid Self-Supervised Learning (ConvMAE + BYOL) pipeline for Brain Tumor MRI classification on 6,000+ images → 97.80% accuracy with Grad-CAM explainability. Paper under review.
• Placement Coordinator & Senior Core — Team Combat Robotics, MIT Manipal
  Operations, recruitment workflows, and full-stack project showcase platform development.

Languages

Backend & Cloud

ML / AI

Frontend & Databases

Winner (Witch Hunt 2026) + 2nd Runner-Up (Codebot IISc)
Event-driven AI surveillance backend with edge filtering. Reduced cloud storage costs by 60% and alert latency by 80% (2.5 min → 30 sec). Later redesigned into a horizontally scalable distributed architecture.
Stack: Node.js, AWS, MongoDB, Event-driven systems

Production-ready offline-first FastAPI service using hybrid rule + DistilBERT routing. Sub-200 ms latency with incremental learning loop (0.84 macro precision).
Stack: FastAPI, DistilBERT, SQLite, Docker

Full-stack platform for dual-income households covering tax optimization, HRA, 80C, emergency funds & FIRE projections with LLM-generated explainable insights.
Stack: Next.js, TypeScript, FastAPI, Python

Strictly client-side meeting assistant with zero middleware. Live audio → Groq Whisper → 120B MoE LLM at 200+ tokens/sec. Absolute data privacy.
Stack: Next.js 15, Web Audio API, Groq, Zustand

Hybrid Self-Supervised Learning (ConvMAE + BYOL) + EfficientNet-B0 pipeline on 6,000+ MRI images → 97.80% accuracy with Grad-CAM explainability.
Stack: PyTorch, EfficientNet, Grad-CAM

Constraint-based invigilator scheduling + role-based academic portal with Keycloak SSO (designed for 120–175+ faculty).
Stack: Spring Boot, PostgreSQL, Keycloak, React

I’m always open to interesting problems in backend systems, applied ML, and low-latency engineering.

Email: work.dhyeypatel@gmail.com
LinkedIn: linkedin.com/in/dhyeypateldev
GitHub: github.com/DHYEYPATL