Skip to content
View UShah1996's full-sized avatar

Highlights

  • Pro

Block or report UShah1996

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
UShah1996/README.md

Hi, I'm Urmi Shah 👋

ML Systems Engineer · LLM Inference · RAG · Production ML
MS Computer Engineering (ML) @ San José State University · 4.0 GPA · Graduating May 2027

LinkedIn Email Open to new grad roles


About me

  • 🏗️ 4+ years in industry at Mercedes-Benz R&D and Tata Consultancy Services, building production ML, data pipelines and LLM services
  • 🔬 Graduate Research Assistant under Prof. Vishnu Pendyala at SJSU, working on transformer internals (previously embodied AI and RL)
  • 👩‍🏫 Graduate Teaching Assistant at SJSU for C++ / Data Structures and for the graduate technical writing course (GWAR)
  • 🩺 Software ML Intern at Cybele Health (Summer 2026): built a full-stack referral platform and an AI profile-enrichment workflow that mixes verified data with LLM suggestions, confidence scoring and human review
  • ⚡ I like the systems side of ML: making models fast, reliable and cheap to serve
  • 🎯 Looking for full-time new grad SWE and ML Engineering roles starting mid-2027

🚀 Featured projects

Project What it does Key results Stack
Weave (MS project, in progress) Runtime-adaptive scheduler that splits LLM layers across 2–3 mixed Apple Silicon Macs and re-partitions live as KV-cache, load and thermals change 3-person team; I own the placement policies, the "how many Macs" decision, and ablation design (adaptive vs. equal, memory-proportional and static splits) MLX / MLX-LM, MLX Distributed, JACCL (RDMA over Thunderbolt 5), Python
Self-Healing Fraud Detection Platform (in progress) Scores card transactions in real time, watches its own accuracy for concept drift, and retrains and hot-swaps the model without downtime 2-person team; I own platform and security (services, streaming, storage, Kubernetes, auth). ~590K real transactions (IEEE-CIS) Microservices, Kubernetes (k3d), Docker, MLflow, S3-compatible storage, token auth
QuadKey: Vacation Rental Booking (in progress) End-to-end booking app with Customer, Host and Admin roles: search, booking, listings, analytics 4-person Scrum team; I own data and infrastructure (database, readiness checks, Docker, AWS deployment) FastAPI, PostgreSQL, React, AWS (RDS, S3, ECR, EC2 Auto Scaling, ALB), GitHub Actions
CardioWatch AFib detection from Apple Watch ECG + clinical data, two-stream late fusion AUC 0.953 ± 0.007 (3-fold CV, 15K ECGs); fixed a hospital-to-wearable domain gap (~50% → 94% on Apple Watch); 30–52 min early warning; AUC 0.909 zero-shot on MIT-BIH Holter PyTorch, XGBoost, SHAP, MLflow, Streamlit · Live demo
HyDE Retrieval Zero-shot dense retrieval with hypothetical document embeddings Recall@10 0.71 with no labeled data, 14.2 ms p99 on 100K MS MARCO passages, 96.4% less memory (307 MB → 11 MB) via product quantization FAISS IVF-PQ, Contriever, Hugging Face
AI-HOPE Multi-agent LLM platform for clinical and genomic data analysis 92% task completion, fully local Llama 2, multi-agent orchestration
FPGA MLP Accelerator Handwritten digit inference on a Zybo Z7-20 INT8 quantized MLP on hardware Zybo Z7-20 FPGA, Python

🔬 Research

  • Empirical Differential Geometry of Transformer Residual Flows: a local flow estimator that cuts held-out update-prediction MSE by 54% (GPT-2 small) and 60% (GPT-2 medium) vs. a global-affine baseline
  • CompGen-EmbodiedAI: compositional generalization study across PPO, hierarchical and Transformer RL agents on a 7-primitive skill benchmark. Zero-shot PPO / hierarchical agents keep 77% / 61% of a depth-3 specialist's success rate
  • Bioinformatics roots: published In Silico Insights into HIV-1 Vpu-Tetherin Interactions and Its Mutational Counterparts (MDPI); poster at BiATA 2020; built genomic data pipelines at CSIR that improved assembly accuracy by 20%

📜 Certifications

DeepLearning.AI Machine Learning Specialization (2026) · Applied Machine Learning in Python, University of Michigan (2026) · SAP Certified Associate, SAP Analytics Cloud: Planning


🛠️ Tech stack

Languages: Python · C++ · C · SQL · Java · TypeScript/JavaScript

ML & LLMs: PyTorch · Hugging Face · vLLM · Apple MLX · FAISS · scikit-learn · XGBoost · MLflow

LLM apps: RAG · HyDE · Agentic pipelines · Azure OpenAI

Data & infra: Apache Spark · Databricks · PostgreSQL · Cassandra · AWS · Azure · Docker · Kubernetes · FastAPI · GitHub Actions

Frontend: React · React Native · Vite


📚 Learning in public

micrograd (autograd from scratch) · makemore-bigram (character-level LM) · cpu-design (CPU design in C++)

Always happy to talk about LLM inference, RAG, and ML systems. Reach out!

Pinned Loading

  1. JNR-10/weave JNR-10/weave Public

    Python 1

  2. cardiowatch cardiowatch Public

    CardioWatch: Early Detection of Atrial Fibrillation Using ECG (CNN-LSTM, AUC=0.968) + Clinical Data (Random Forest, XGBoost) — Apple Watch compatible

    Jupyter Notebook

  3. HyDE HyDE Public

    Minimal HyDE: Zero-Shot Dense Retrieval with FAISS IVF-PQ

    Python

  4. AI-HOPE AI-HOPE Public

    AI-HOPE: Multi-Agent LLM Platform for Clinical & Genomic Data Analysis

    Python

  5. 240_mlp_fpga 240_mlp_fpga Public

    FPGA-Based MLP Accelerator for Handwritten Digit Recognition

    Python