Independent Researcher | AI Research Engineer @ VFS Global
About
Hi, I’m Satwik. AI Research Engineer at VFS Global. Independently researching reliability, reasoning, and interpretability of language models.
At VFS Global, I work on production document intelligence for multi-country visa processing. I build extraction and verification systems for real document workflows, including multimodal validation, confidence-aware routing, and checks around model outputs. These systems sit in workflows that process 4M+ documents daily.
My research is on reasoning and agentic systems, with a focus on making them safe, reliable, and uncertainty-aware. Much of that comes down to calibration and failure detection, and to whether these systems hold up once they're deployed rather than just benchmarked.
Previously, I worked on trustworthy reasoning for LLMs at UCSC’s AIEA Lab. During undergrad, I also spent roughly two years across research, open-source, and engineering internships, mostly spanning server-side systems and applied AI pipelines.
News
Two papers accepted at the Failure Modes in Agentic AI (FAGEN) workshop at ICML 2026: SELFDOUBT and Proper Scoring Rules for Agentic Uncertainty Quantification.
Proper Scoring Rules for Agentic Uncertainty Quantification accepted as a poster at the Combining Theory and Benchmarks (CTB) workshop at ICML 2026.
New preprint: Proper Scoring Rules for Agentic Uncertainty Quantification. We introduce the Trajectory Proper Score (TPS), a predictor-agnostic family of strictly proper, trajectory-level scoring rules for evaluating uncertainty in LLM agents.
New preprint: SELFDOUBT: Uncertainty Quantification for Reasoning LLMs via the Hedge-to-Verify Ratio. We introduce HVR, a single-pass uncertainty signal for reasoning LLMs that outperforms Semantic Entropy at about 10x lower inference cost.
New preprint: Don’t Blink: Evidence Collapse during Multimodal Reasoning. We identify evidence collapse, a decay of visual grounding during multimodal reasoning that text-only uncertainty signals cannot detect.
Joined VFS Global as an AI Research Engineer.
Publications & Preprints
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 are not strictly proper in the agentic setting.
Proposed an O(1) black-box uncertainty framework that extracts behavioral hedge/verify signals from reasoning traces, significantly outperforming Semantic Entropy on discrimination (p=0.001) at 10× lower cost; a zero-hedge gate achieves 96.1% precision across 7 models and 3 benchmarks.
Identified a universal evidence collapse phenomenon in reasoning VLMs, observing visual attention drops up to 90.8% during generation and a task-conditional failure regime where confident but visually disengaged predictions are hazardous on sustained visual reference tasks but benign on symbolic tasks.
Introduced a function-level APR framework achieving a SOTA 83.1% plausible repair rate on Defects4J, with an automated verification pipeline combining AST alignment, control-flow symbolic analysis, and semantic checks.