CV

AI scientist and research engineer working on agentic systems, world models and learning systems. My work spans research (ICAPS, IJCAI, one of the named inventors on a pending patent) and production AI platforms.

[email protected] · LinkedIn · GitHub

Experience

Senior Member of Technical Staff, AI

Didero, New York · Jan 2025 to present

  • First on the team to push for an agentic architecture for customer-specific procurement workflows, replacing fixed pipelines that couldn’t handle per-customer variation.
  • Architected the core agent platform that enabled Didero’s Series A and the onboarding of new enterprise customers. It adapts to each customer’s workflow through configuration rather than a rebuild.
  • Built and drove a closed-loop, self-improving agent platform that replaced hand-written evals, previously a multi-day bottleneck for every new customer.
  • Built the in-house agent tracing system used for debugging, eval generation and as the input signal for optimization.
  • Built a reflection-based prompt optimization system with a feedback analyzer that clusters production rejections into structured signals.
  • Extended the platform with a continuous-learning loop, leading the team’s implementation while building core pieces directly.
  • Built a freight-invoice audit agent at an internal hackathon, chosen as the top project and later productized.
  • Started the web-browsing agents that integrate with ERPs and other systems that have no API.

AI Research Scientist

Andiron AI, Washington, D.C. · Dec 2023 to Dec 2024

  • Built a RAG and web-search multi-LLM framework for deep market research and content generation, months before OpenAI’s Deep Research.
  • Prototyped a Voyager-inspired multi-agent system with AutoGen for self-directed SEO optimization.
  • Cut a keyword research pipeline from 30 minutes and 128 GB of RAM to on-demand runs under 4 GB.

Deep Learning Research Engineer

Geopipe (acquired by NBCUniversal), New York · Apr 2021 to Oct 2023

  • One of the named inventors on a pending patent for the Neural Building Painter, Geopipe’s ViT-based second-generation model for 3D reconstruction from aerial imagery and LiDAR.
  • Worked on the Neural Building Printer (first generation, Pix2PixHD-based GAN, issued patent), improving output from rough reconstructions to usable 3D geometry.
  • Extended 2D IoU into a 3D IoU loss that improved training stability and reconstruction accuracy.
  • Sped up preprocessing roughly 120x by vectorizing matrix operations.
  • Joined as the fourth full-time hire and worked closely with the Chief Science Officer on research direction and infrastructure.

Graduate Research Assistant

University of Georgia, THINC Lab · Aug 2018 to Apr 2021

  • Invented Recurrent Sum-Product-Max Networks (RSPMNs) for sequential decision-making and offline reinforcement learning, and implemented them end to end.
  • Outperformed contemporary deep-learning approaches on sequential decision-making benchmarks across 7 custom RL environments.
  • Published at ICAPS (first author) and IJCAI. Open-source code drew interest from peer labs, including a group at MIT.
  • Contributed to securing NSF funding through grant writing and experimental planning.

Systems Engineer

Amex, TCS, Bengaluru · May 2015 to May 2017

Research and Projects

World models for robot control (Carnegie Mellon, 2026). Led our team’s world-models research for language-conditioned robot arm control on VLA-Arena. Found that DINOv2, V-JEPA and Cosmos backbones all converged to the same ceiling, and that in low-data regimes world models work better as policy conditioners than as planners.

Web navigation with a small LLM (Carnegie Mellon, 2025). Fine-tuned LLaMA 3.2 (3B) for browser navigation from GPT-4.1 tool-call traces, taking it from 0% to reliable basic navigation. Model

LLM training infrastructure and GPU kernels (Carnegie Mellon, 2025). Built MiniTorch (tensors, autograd, training loops), wrote CUDA kernels for matrix operations and attention, and implemented LLM pretraining and evaluation pipelines.

Publications and Patents

  • Data-Driven Decision-Theoretic Planning using RSPMNs. ICAPS 2021. Paper
  • State-Based RSPMNs for Partially-Observable Sequential Decision Making. IJCAI 2021.
  • Neural Building Painter (ViT-based 3D reconstruction), Geopipe / NBCUniversal. Patent pending, one of the named inventors.
  • Neural Building Printer (GAN-based 3D reconstruction), Geopipe / NBCUniversal. Issued patent, contributed after filing.

Education

  • Graduate Specialization, Generative AI and LLM Systems. Carnegie Mellon University, 2025 to 2026.
  • M.S., Artificial Intelligence. University of Georgia, 2017 to 2020. ROOST Scholarship.
  • B.Tech., Electrical and Electronics Engineering. SASTRA University, India, 2011 to 2015.

Skills

  • Areas: LLMs, agentic systems, reinforcement learning, world models, multimodal learning
  • ML: PyTorch, TensorFlow, JAX, Hugging Face, Unsloth
  • Agents: Pydantic-AI, AutoGen, CrewAI, LangGraph, DSPy, Browser-Use
  • Systems: Python, C++, CUDA, Docker, AWS, GCP, Temporal, Langfuse

Service

  • Technical writing at this site, 2025 to present
  • AI mentor, AI4All, 2022
  • Organizer, ACM NYC, 2022 to 2023
  • Teaching volunteer for primary school students in India, 2015 to 2017