cv

B.Tech

Vellore Institute of Technology, Bhopal

Computer Science and Engineering

CGPA: 3.6/4.0

B.Sc.

Indian Institute of Technology, Madras

Programming & Data Science

CGPA: 3.3/4.0, Project CGPA: 4.0/4.0

Don't Blink: Evidence Collapse during Multimodal Reasoning

Suresh Raghu*, Satwik Pandey* | Under Review | * Equal contribution

Identified "evidence collapse", a universal (9/9 model x dataset cells) decay of visual grounding during VLM reasoning that text-only entropy cannot detect, and designed a task-conditional vision veto that cuts selective risk by up to 1.9 pp at 90% coverage on MathVista, HallusionBench, and MMMU_Pro.

SELFDOUBT: Uncertainty Quantification for Reasoning LLMs via the Hedge-to-Verify Ratio

Satwik Pandey*, Suresh Raghu*, Shashwat Pandey | Under Review | Accepted (poster) at FAGEN Workshop @ ICML 2026

Proposed the Hedge-to-Verify Ratio (HVR), a single-pass O(1) uncertainty signal that scores a reasoning trace by how much it hedges versus self-verifies. Traces with no hedging language are correct 96.1% of the time (at 25.4% coverage), and fusing HVR with the model's verbalized confidence beats sampling-based Semantic Entropy on discrimination (p=0.001) at 10x lower inference cost, across 7 models and 3 benchmarks.

Proper Scoring Rules for Agentic Uncertainty Quantification

Suresh Raghu*, Satwik Pandey*, Shashwat Pandey | Under Review | Accepted (poster) at CTB & FAGEN Workshops @ ICML 2026

Introduced a family of strictly proper trajectory-level scoring rules for evaluating uncertainty in LM agents, with a censored-trace extension and negative results showing that standard trajectory-level calibration metrics, including ECE and Brier, are not strictly proper in the agentic setting.

Manager -- AI (Founding Lead, AI Research Engineering)

VFS Global

Leading an AI team building production intelligent document processing systems that process 180K+ documents daily across 110+ countries.

• Adaptive Query Routing: Architected the document-extraction pipeline around an adaptive router that sends deterministic documents to lightweight parsers and ambiguous ones to reasoning VLMs, reaching 99.1% field-level extraction accuracy and reducing visa review time from 60+ minutes to sub-5 minutes.
• Uncertainty-Gated Extraction Cascade: Used a cheap token-logit trigger to run Semantic Entropy and Monte Carlo self-consistency only on suspect outputs, distinguishing linguistic variance from epistemic model failure and escalating flagged cases to a larger LLM for dual-model validation, cutting critical extraction errors by 87%.
• Biometric Compliance Pipeline: Proposed and built a two-stage visa photo-verification system consisting of a cheap image-quality gate (Laplacian blur detection) followed by a VLM stage for pose/occlusion checks, reducing cases requiring manual intervention by 78%.
• Technical Leadership: Led a technical team of 8+ engineers.

Applied Machine Learning Intern

Adani Digital Labs

Developed ML systems for content automation, computer vision, and multimodal search.

• Content Automation: Delivered an OpenAI API pipeline generating SEO-optimized content across 150K+ pages, achieving $3.7M in cost savings.
• Self-Supervised Representation Learning: Used Meta's DINOv2 embeddings with an SVM over latent-space projections for few-shot image clustering on a 500+ image dataset, replacing manual tagging at 93% precision.
• Multimodal Search: Built a CLIP + GPT-4 Vision product-discovery engine supporting both text and image queries over proprietary retail inventory.