I build systems that turn messy data and fuzzy ideas into things that ship: from LLM research pipelines to production APIs.
AI/ML Engineer & Full-Stack Developer · Los Angeles, CA
About
A little about me.
I'm a Computer Science graduate of Cal State LA (GPA 4.0), where I worked as a Graduate Research Assistant on ATA-LLM, a framework that turns thousands of raw interview transcripts into coherent themes using LLM embeddings and unsupervised clustering.
Outside research, I build full-stack and agentic systems: from a news platform with AI summarization to an RPA platform that eliminates repeat LLM costs entirely. I care about work that's validated with real numbers, not vibes: statistical significance tests, benchmark comparisons, PyTest coverage.
Based in Los Angeles, originally from Gujarat, India.
Education
Where it was built.
California State University, Los Angeles
08/2024 – 05/2026Master of Science, Computer Science
GPA 4.0 / 4.0
Charotar University of Science and Technology, India
10/2020 – 05/2024Bachelor of Technology, Information Technology
GPA 3.92 / 4.0
Skills
The stack, mapped out.
Experience
What I've been building.
Graduate Research Assistant
09/2024 – 05/2026
California State University, Los Angeles
- Processed 4,536 open codes from 175 interview transcripts by developing ATA-LLM, a framework combining 1,536-dimensional LLM embeddings, UMAP/DenseMAP dimensionality reduction, and HDBSCAN clustering, with full PyTest coverage.
- Achieved 0.893 topic coherence and 77% cosine-similarity alignment with human-coded themes across 200+ Optuna Bayesian optimization trials, outperforming t-SNE baselines (0.817 TC) and confirming results via statistical significance testing.
Data Science Intern
01/2024 – 05/2024
Collabera Digital, Gujarat, India
- Delivered 83% accuracy in automated skill extraction by fine-tuning and deploying BERT-based NLP models (PyTest-covered) via production REST APIs, reducing manual screening effort by 40% across large-scale recruitment workflows.
- Improved candidate-job matching precision by 25%, validated through A/B and statistical significance testing, by building GPT-based prompt engineering pipelines and curating labeled training datasets for scalable, automated talent mapping.
Android Developer Intern
05/2023 – 07/2023
iTeam Technology, Gujarat, India
- Reduced backend response latency by 35% by building and verifying production-ready Android features using Java, JUnit/Espresso, RESTful APIs, and MySQL, enabling real-time data interactions for concurrent users.
- Strengthened data consistency by 20% by designing normalized relational schemas and optimizing SQL queries to support real-time inventory tracking and transaction processing.
Projects
Things I've shipped.
AI-Powered News Aggregation Platform
Neuz Now
A role-based news platform that uses generative AI to auto-summarize articles, cutting reading time by more than half.
- Enabled secure, role-based access control across 3 user roles (Reader, Author, Admin) by architecting a news aggregation platform using React Native, Node.js, and MySQL with JWT authentication and RBAC.
- Decreased reading time by 60% by integrating Google Generative AI into a Node.js/Express.js pipeline to auto-summarize news articles, boosting engagement through personalized content delivery.
LLM-driven discovery with a deterministic replay engine
Agentic RPA Platform
An agentic automation system that separates expensive LLM-driven discovery from a free, deterministic replay engine, cutting repeat-run cost to zero.
- Eliminated 100% of LLM inference cost and latency on repeat runs by architecting an agentic system that separates LLM-driven discovery from a deterministic, model-free replay engine, reused across tenants via a versioned override layer.
- Enabled safe operation on an IT-helpdesk console via a policy engine with allowlisting, risk classification, and PII redaction, plus an agent-facing catalog with a stability gate catching 2 real bugs, validated by 35 tests.
Self-attention for 3D vision
Point Cloud Processing with Point Transformer
A Point Transformer with a custom offset-attention U-Net decoder, benchmarked against PointNet and voxel/projection-based baselines on ModelNet10.
- Achieved 92.40% classification accuracy (3.46% loss) on the ModelNet10 benchmark (4,899 shapes, 10 categories) by implementing a Point Transformer with self-attention and positional encoding.
- Optimized spatial feature retention across scales by engineering a 4-layer offset-attention U-Net decoder with 5 downsampling and 5 upsampling stages, benchmarked against PointNet, voxel-based, and projection-based architectures.
Publications
Peer-reviewed work.
Exploring Manifold-Based Clustering Techniques for Enhanced Inductive Thematic Analysis
Springer Nature · 2025
Comprehensive Study and Analysis of Point Transformer for Point Cloud Data
Journal of Propulsion Technology · 2024
Number Plate Detection and Recognition using OpenCV
IEEE Xplore · 2024