Paras Goel
I'm a computer science and applied mathematics student at Columbia University interested in building intelligent systems at the intersection of AI agents, generative models, and systems engineering. Currently working on Copilot Frontier Tuning at Microsoft and video generation at DitecT Lab.
What I Work With
Software Development Intern, Copilot Tuning Team
- ▹ Developed evaluation and experimentation infrastructure for the Copilot Frontier Tuning (FT) platform, enabling reinforcement learning based optimization of enterprise AI agents and LLMs.
- ▹ Built & scaled performance analytics for 1,000+ agents across 250+ FT tenants, used daily by 30+ developers.
- ▹ Leveraging Monte Carlo Tree Search to evaluate FT models' reasoning, latency, and token efficiency tradeoffs.
- ▹ Developing an autonomous AI agent to orchestrate feature rollout workflows across 33,000+ flights, including tenant onboarding, deployment debugging, & configuration validation.
ML Researcher
- ▹ Developing causal inference frameworks for video generation by modeling temporal dependencies and intervention effects, enabling more controllable and interpretable generation dynamics.
- ▹ Adapting causal forcing techniques to distill large bidirectional generative models into efficient autoregressive architectures, reducing inference cost while preserving generation quality.
- ▹ Analyzing temporal consistency, causal fidelity, and generation efficiency across video synthesis models.
AI Engineer
- ▹ Developed OAuth-based authorization workflow for enterprise AI agents, enabling secure access to customer resources while maintaining isolation across agent workflows.
- ▹ Designed and deployed MCP servers and tool-integration layers supporting authenticated communication between LLM agents, external services, and internal data systems.
Software Developer
- ▹ Leveraged LLMs to transform unstructured business content into production-ready storefronts, enabling rapid merchant onboarding with generated catalogs & integrated payments (Alpha tested with 100+ customers).
- ▹ Developed infrastructure for full-stack system by integrating multimodal image generation models, LLMs, and PayPal SDKs with responsive UIs and scalable cloud deployment on Vercel.
- ▹ Created evaluation pipelines for AI models' outputs for catalog creation, pricing, natural language-to-HTML conversion, and price accuracy.
AI/ML Researcher
- ▹ Implemented 3D attention mechanisms for glioma classification from multimodal MRI scans.
- ▹ Built image-to-image deep learning pipeline using latent diffusion model for Region of Interest extraction.
Utilizing MCP and Shared Introspection for Targeted Agent-to-Agent Communication in Hierarchical Clinical Multi-Agent Systems
We developed a hierarchical multi-agent system for clinical decision support, combining five domain-specialist agents with an attending-style meta-agent to coordinate complex medical reasoning. Through structured deliberation and confidence-based conflict resolution, our system improved diagnostic accuracy from 20% to 80% on USMLE-style clinical questions while reducing response time by approximately 28%.
Facilitating Shared Introspection in Hierarchical Multi-Agent Systems Using the Model Context Protocol
We developed an MCP-enabled multi-agent framework that allows specialized medical AI agents to share and reflect on their reasoning in real time. By coordinating these insights through a supervising agent, we explored how structured introspection and deliberation can make clinical AI more transparent, efficient, and resilient to individual agent errors.
Utilizing Multi-Agent Reinforcement Learning with Encoder-Decoder Architecture Agents to Identify Optimal Resection Location in Glioblastoma Multiforme Patients
We developed an end-to-end AI framework for glioblastoma diagnosis and treatment planning, combining sequential classification with generative models that simulate surgical resection, radiotherapy, and chemotherapy outcomes. By integrating these models with reinforcement learning, our system iteratively explores treatment strategies toward a desired survival outcome while substantially reducing diagnostic compute costs and tumor-progression inference time.
CoCa-Mil: Attention-Based Handcrafted-Deep Feature Fusion in Computational Pathology
We developed CoCa-MIL, an attention-based framework for whole-slide image classification that combines domain-specific handcrafted features with representations learned by deep neural networks. Using co-attention and cross-attention mechanisms, our approach leverages the complementary strengths of both feature types, improving classification accuracy on the TCGA Lung Cancer dataset by up to 5.21% over baseline methods.
CA-fuse-MIL: cross-attention fusion of handcrafted and deep features for whole slide image classification
We introduced CA-Fuse-MIL, a cross-attention approach that integrates handcrafted pathology features with deep learned representations for whole-slide image classification. We also explored multi-layer variants of the architecture, showing that explicitly learning interactions between these complementary feature spaces can outperform deep features alone, with accuracy gains of up to 5.21% on the TCGA Lung Cancer dataset.
Role of stain normalization in computational pathology: use case in metastatic tissue classification
We investigated how stain normalization and color augmentation can improve the robustness of deep learning models across variations in histopathology imaging. Evaluating six preprocessing configurations on more than 300,000 tissue images, we found that combining Macenko normalization with color augmentation produced the strongest metastatic tissue classification performance, improving both accuracy and F1 score over the baseline.
Education & Involvements
Columbia University
Generative models, multi-agent reasoning, causal inference, and building production-scale AI architectures.