About
My first taste of machine learning was open-source work at IIT Bombayin 2018. After that I went into industry: Stylumia for multimodal search,Byju's for transformers in production, and G2for two years of graph neural networks over user-product data.
I finished my Master's in Computer Science at UT Dallas in December 2025. Alongside it I interned at ISN on multi-agent systems and a multimodal risk advisor, and I still do research at UT Dallas on adversarial red-teaming agents for network protocols.
Today I'm a machine learning engineer at a major US telecom, building agents that resolve network faults end to end. Before that, and for a few months alongside it, I was a part-time founding engineer on a VR learning platform.
On the side I contribute to OpenClimateFix, an open-source group forecasting solar power, and I run my own GPU server at home for local models.
Recently working with: Graph Neural Networks · Natural Language Processing · PyTorch · LangChain & LangGraph · Azure and AWS · Terraform and CI/CD · MLflow and Kubeflow · llama.cpp & local LLMs.
Experience
Machine Learning Engineer · A major US telecom
- Built the agentic orchestration behind automated fault resolution for business fiber across the US customer base, from prediction through dispatch, field follow-up and ticket closure. Outage resolution went from days to hours
- Put deterministic guardrails around LLM classification, which cut false positives by roughly 75%
- Connected agent workflows to 18 vendor platforms with strictly typed Python contracts
- Built a real-time lead ingestion and propensity-scoring pipeline (doubly robust estimation, inverse probability weighting) with voice agents for inbound and outbound engagement. It finds sales that would otherwise have been missed
Founding Engineer (part-time) · Stealth startup, AI-powered VR learning
- Held about 90 ms bidirectional speech latency across 10,000+ daily messages on Azure Functions
- Built the backend on Django and PostgreSQL: 15+ REST endpoints with organization-scoped RBAC, bulk assignment and multi-tenant indexing, sustaining about 1,000 peak concurrent users
- Designed the multi-tenant SaaS infrastructure on Azure (6 Terraform components, CI/CD, OIDC, zero-downtime deploys), piloted with a fast-food chain and a bank with 1,000+ employees
Data Science Intern · ISN
- Designed and deployed a multi-agent solution on Azure, including the serving infrastructure for a customer-facing chat agent and an internal copilot. It ingests 10+ TB of call recordings
- Led a 0-to-1 multimodal Risk Advisor from concept to deployment for 500,000+ industrial workers, processing text, video and geospatial data in real time
- Built facial ID verification into ISN’s Empower app, serving 500,000+ users
Graduate Research Assistant · UT Dallas
- Building an adversarial red-teaming agent for network protocols. An LLM proposes multi-hop attack sequences, and a live testbed runs them (12k unique sequences so far)
- Closed a reinforcement loop from the testbed back into the model: LoRA fine-tuned a 35B-parameter mixture-of-experts model across 2 H200 GPUs with DDP, lifting valid query generation from 0% to 39% on held-out prompts
- Finished an MS in Computer Science in December 2025
Data Scientist II · G2
- Built and maintained deployment pipelines for user-product interaction models, with model versioning and experimentation, in a production system serving 90 million users a year
- Built and scaled graph neural networks over 20+ TB of user-product interaction data, with vector databases and graph representation learning pipelines behind recommendations and downstream systems
- Scaled a multimodal auto-moderation pipeline that saved 40% of manual effort at 90%+ confidence
Research Engineer · BYJU'S
- Optimized transformer architectures in production, cutting response time by 30% and lifting throughput by 15%
- Built a solution for math word problems by deploying GNN and NLP models into production
- Used structured paraphrasing to widen the diversity and quality of the training set
Data Scientist · Stylumia
- Built a multimodal vertical search engine (NLU, BERT-as-a-service) that improved search relevance by 68% and cut response time by 0.3 s. It sits on a hierarchical ontology that organizes 4 million products into 10,000+ brands and categories
Research Scholar · IIT Bombay
- Contributed to openEdX: a peer-assessment module, automated grading and decentralized content management
Contact
Looking for a data scientist, want to argue about graphs, or found something broken on this site?Email me; I actually reply.