Satwik Pandey

Mohamed bin Zayed University of Artificial Intelligence · Kang Lab

Research Collaborator · Advisor: Prof. Jian Kang

• Investigating decorative verification in reasoning LMs: whether self-checking language such as verification, backtracking, and reconsideration can occur without the causal computation needed to update model beliefs.
• Studying which reasoning operations—including alternative exploration, contradiction detection, backtracking, and evidence integration—genuinely improve calibration versus merely signal epistemic caution.

University of California, Santa Cruz · AIEA Lab

Research Assistant · Advisor: Prof. Leilani H. Gilpin

• Investigated trustworthy logical reasoning in LLMs by translating natural language outputs into Prolog-based logic constraints to mathematically verify consistency and detect hallucinations.
• Designed a Modular RAG architecture to decouple knowledge retrieval from reasoning, reducing false positive hallucination detection by integrating external ground-truth validation.

Jaypee Institute of Information Technology

Undergraduate Researcher · Advisor: Prof. Sandeep Kumar Singh

• Conducted initial experiments on parameter-efficient fine-tuning (Unsloth/QLoRA) which revealed limitations in complex logic repair, motivating a strategic pivot from training-based optimization to an inference-time reasoning framework that became the core contribution of the Repair of Thought publication.

Jaypee Institute of Information Technology

Bachelor of Technology in Computer Science · CGPA: 3.3 / 4.0

VFS Global

AI Research Engineer

• Investigated dynamic execution paths for multi-agent frameworks. Designed a complexity-based routing mechanism that differentiates between deterministic tasks handled by parsers and ambiguous instances routed to reasoning-capable VLMs, optimizing the trade-off between inference cost and reasoning accuracy.
• Integrated a two-stage confidence pipeline: token-level logit probabilities aggregated via Shannon entropy serve as a lightweight first-pass filter for extraction uncertainty, with high-entropy predictions escalated to a reasoning VLM-as-a-Judge for cross-modal extraction verification.
• Designed a trace-guided re-extraction mechanism for cascade-rejected fields that fail both entropy screening and judge validation. The judge's reasoning traces are converted into corrective context for a secondary extraction model, enabling targeted recovery from extraction errors such as character misreads, boundary confusion, and semantic misattribution.
• Designed a multi-level verification framework for biometric photo validation, separating signal-level quality assessment from semantic-level attribute verification with fine-tuned VLMs. The staged design isolated distinct failure modes and improved the interpretability of rejection decisions.