AI/ML Thought Leader | Data Storyteller | Mentor | Inventor | Senior Member IEEE
🚀 Extremely proud and excited to share that the Centific AI Research team has three papers accepted at NeurIPS 2025! Each tackles a critical problem at the intersection of AI governance, trust & safety, and data collection and annotations for language, vision, and world models. 1️⃣ LegalWiz: Multi-Agent Contradiction Detection in Legal Documents https://lnkd.in/gkdj5BfS **Why it matters:** Legal and policy documents often contain subtle inconsistencies, leading to serious compliance and contractual risks. Existing contradiction benchmarks are too simplistic for these high-stakes contexts. **What we built:** LegalWiz is a multi-agent generation framework that automatically injects, detects, and verifies contradictions across legal-style documents—combining NLI models with LLM judges in a human-in-the-loop loop. The hybrid system improves precision and recall by over 20% versus standalone models, offering one of the first realistic contradiction-aware benchmarks for RAG pipelines in law and governance. 2️⃣ RECAP: Hybrid Methods for Multilingual PII Detection https://lnkd.in/gbZjqMnj **Why it matters:** Global privacy compliance depends on reliable PII detection—but low-resource languages remain a blind spot due to lack of training data. **Our approach:** RECAP fuses regex-based precision with LLM-based semantic reasoning in a modular, locale-aware architecture covering 13 languages and 300+ PII entity types. Its three-phase refinement pipeline (disambiguation → consolidation → contextual filtering) boosts weighted F1-score by 82% over NER and 17% over zero-shot LLMs, offering a scalable privacy safeguard for multilingual data operations. 3️⃣ GAZE: Governance-Aware Pre-Annotation for Zero-Shot World Model Environments **Why it matters:** Training large-scale world models needs massive multimodal video data—but annotating hours of footage is expensive, inconsistent, and privacy-sensitive. **What we introduced:** GAZE is a governance-first pre-annotation pipeline that transforms raw 360° video into structured, privacy-aware supervision. It integrates multi-task vision, audio, and governance filters (PII, minors, NSFW, motion, ASR, scene understanding) and enables review-by-exception workflows, cutting human review effort by ~30% (≈19 min/hour saved). The result: trusted, legally-compliant world model datasets at production scale. These works advance trustworthy AI, data governance, multimodal intelligence, and World Models — moving toward safer, auditable AI pipelines across domains. Congratulations to Vasudevan ( Vasu ) Sundarababu, Leela K., Harshit Rajgarhia, Ananya Mantravadi, Shivali Dalmia, Sai Charith Reddy Pasula, Suryam Gupta, Asif Shaik, Santhoshraj Y, Nithya Tanvi Nishitha S, PraveenKumar Gulipalli, and other contributors. We look forward to discussing cutting-edge research with fellow researchers. #NeurIPS2025 #ResponsibleAI #AIResearch #Governance #LLM #MultimodalAI #Centific #AICompliance #WorldModels
Congratulations to the Centific AI Research team on these outstanding NeurIPS 2025 acceptances! Each paper tackles such timely and essential challenges in AI governance and safety — amazing to see this level of innovation. 👏
Great work team!
Ankit S.
3w
Congratulations
Awesome Abhishek Mukherji, Ph.D., way to go!
Shipping production controls for reliable AI agents → coa.dev. Data and AI Product builder with a history of working with large scale distributed systems for analytics, AI/ML.
2wCongratulations to the Centific AI Research team! These papers are an excellent example of advancing trustworthy AI and governance at scale. LegalWiz, RECAP, and GAZE address real-world risks, from legal contradictions to multilingual PII detection and privacy-aware world model annotation showing the importance of auditable, multi-agent pipelines. For enterprises deploying similar AI workflows, platforms like CoAgent can help monitor agentic systems, validate outputs, and ensure compliance across complex pipelines. Observability combined with automated evaluation makes it possible to scale these innovations safely and reliably. It would be interesting to hear how others are operationalizing governance-first AI while maintaining flexibility for model updates and multi-modal workflows.