Build fast.
Learn faster.
MS in Computer Science, Binghamton. Ships agentic AI systems end to end, with the retrieval, ranking and test coverage to back them up.
Roshni is an AI/ML software engineer in the Bay Area working across Python, C++ and TypeScript: FastAPI and GraphQL backends, React frontends, RAG and hybrid vector search, LangGraph agents, and pytest/Playwright suites that run against Docker Compose. She built an agentic career-planning system for Binghamton’s Watson Career Services, cut document-processing time 40% at Kanini, and keeps winning at San Francisco hackathons on weekends. Debugs to root cause. Leaves tests behind.
in See the full picture on LinkedIn — 500+ connections, hackathon recaps, photos →Languages
- Python
- C / C++
- JavaScript
- TypeScript
- SQL
Software engineering
- Data structures & algorithms
- REST & GraphQL APIs
- FastAPI, Django, Flask
- React
- Root cause analysis
Testing & infra
- pytest, Playwright, Jest
- Smoke & end-to-end testing
- Docker, Kubernetes
- AWS EC2, Linux
- Git
Generative AI & data
- LLMs & multi-modal models
- RAG, LangGraph, MCP
- PyTorch
- PostgreSQL, MongoDB
- ClickHouse, Neo4j
Graduate Assistant
Binghamton University, SUNY
- Architected an agentic AI system for career planning and role matching: skill-gap analysis and personalized learning roadmaps, with prompt engineering for role-fit explanations.
- Partnered with Watson Career Services to gather requirements, test workflows and iterate on agent behavior from stakeholder feedback ahead of a planned release for student use.
- The work grew into EvidPath, the full-stack career-planning platform.
Data Analyst Intern
Kanini
- Shipped a full-stack document-analysis application with Django and Flask backed by PyTorch NLP pipelines, cutting manual processing time by 40%.
- Built and deployed a document classification model pairing an SVM classifier with BERT embeddings, served through Flask APIs; also used OpenAI and Transformers embeddings.
- Exposed REST APIs across the application for modular, reusable deployment, and built Python data pipelines and stakeholder dashboards.
AI/ML Intern
University of Texas at Dallas
- Worked with a cross-functional research team on Python solutions for classifying and analyzing multimedia data.
- Trained a music-genre classifier with K-Nearest Neighbors and scikit-learn, reaching 82% on GTZAN after hyperparameter tuning.
- Implemented the MFCC feature-extraction pipeline (Librosa, NumPy) and the model-evaluation pipeline.
MS Computer Science
State University of New York at Binghamton, Watson College. Algorithms, Systems Programming, Mobile System Security, AI, Data Mining.
B.Tech Information Technology
Rajalakshmi Engineering College, Chennai. GPA 8.62/10. DSA, Networks, DBMS, Operating Systems, Software Testing.
CoreWeave Hacks: Agent Loops
We built Arian OS, agents that read a published ML paper and keep experimenting until a chosen metric improves. Two awards at one event, September 2026.
CASE STUDYBetter Days Hackathon
We built an early warning system that tracks how quickly river levels rise instead of waiting for rain. ClickHouse and LibreChat hosted it in San Francisco, August 2026.
CASE STUDYAI Infra Summit Hackathon
We built PEMA, an on-device assistant that remembers where objects were last seen for people with low vision or memory loss. Santa Clara, September 2026.
CASE STUDYTwelve-plus hackathons so far, mostly in San Francisco. The rest of them, with photos, are on LinkedIn.
Oracle Agentic AI Foundation
Oracle
Women in Tech Salesforce Training Program
MTX
Robotic Process Automation
UiPath
Machine Learning & Deep Learning
The University of Texas at Dallas
Python for Data Science
Great Lakes Learning Academy
Selected
Works.
EvidPath
AI career-planning platform: RAG resume analysis and skill-gap scoring across 205 role pathways.
Silent Flood
Detects a flood by how fast the river is rising, not by whether it is raining.
Arian OS
Agents that read a published ML paper and keep iterating until your metric improves.
VariantCourt
An AI courtroom for DNA variant interpretation, where a rule engine — not an LLM — gives the verdict.
Successor
Voice-based knowledge capture for semiconductor fab operations.
CivicTrace SF
Government meeting video → a Neo4j knowledge graph that catches contradictions.
PEMA
Personal Environmental Memory Agent — Physical AI that remembers where things are.
EchoLoop
You don't learn how to talk to the AI. The AI learns how you communicate.
CloudVault Lite
A deduplicating, content-addressable storage engine in C++.
ClearForm + Readback
Voice-first document completion for blind and low-vision users.
ATE Yield Tracker
ATE log analysis and PCB yield tracking for manufacturing test engineering.
CCTV Anomaly Detection
An LSTM pipeline that classifies shoplifting, robbery and explosions in short CCTV clips and files an incident report.
Warzone Tweet Sentiment
Multi-model sentiment classifier tracking public opinion on the Russia–Ukraine conflict at 94% accuracy.
Flipkart Data Bot
An unattended UiPath RPA bot that extracts structured product data from Flipkart into Excel.
Everything else, including the smaller builds, is on GitHub.
Let’s
Talk.
Hiring for AI engineering, backend or ML? Email me, my inbox is open.