RONOK
TANVIR

Real-time fraud detection system with an IsolationForest ML model hitting 88% precision, backed by SHAP-based interpretability so every flagged transaction comes with a human-readable explanation. Built in 36 hrs at Cal Hacks 12.0, secured the BrightData sponsorship track, and is in active product development with the BrightData team.

Multi-agent RL environment where 7 LLM agents play Diplomacy — negotiating, deceiving, and betraying — while a separately trained overseer model learns to infer each agent's true strategic intent from behavioral signals alone. The overseer never sees private messages; it reads actions, conflict patterns, and communication metadata to predict what each agent is actually planning. Trained via GRPO with an LLM judge as the reward signal. The headline result: a 52.1% judge agreement rate from a 0% baseline — detecting betrayal before it happens in adversarial AI systems.

Multimodal video understanding pipeline that uses TwelveLabs Pegasus and Marengo models to transform long-form educational content into structured, digestible learning modules. Async processing via Django, Celery, and Redis keeps it fast at scale. Selected to present at a TwelveLabs developer webinar secured through cold outreach.

Computer vision model for identifying AI-generated and manipulated media using a CNN-based architecture trained on real vs. synthetic image datasets. Designed as a modular, plug-in component for authenticity verification pipelines, with a focus on generalizing across generation methods rather than overfitting to a single GAN or diffusion model.