Built an OpenEnv-compliant RL gym that trains Qwen2.5-1.5B + LoRA as a prompt-rewriting safety layer for diffusion models, with a 3-way action space (allow / transform / reject). GRPO training improved reward from −0.05 → +0.33 over 1,300 steps with zero mode collapse. Designed a 4-term reward combining Llama Guard 3, a CLIP concept-arithmetic intent residual, a NudeNet/Q16 visual judge, and an over-refusal penalty. Top 100 of 5,000+ teams at the Meta × OpenEnv Hackathon 2026.

Engineered a DistilBERT safety classifier (87% F1 across 3 classes) with batch inference and confidence scoring, plus an end-to-end red-teaming pipeline (500+ adversarial prompts, 3 harm categories, 3 mitigation agents) that reduced unsafe generations by 42%. Built an observability suite - confusion matrices, per-class metrics, and JSON/visual exports - for real-time monitoring and drift detection in production LLM safety workflows.

Translating 2-band SAR imagery into 13-band multispectral EO images using CycleGAN while preserving semantic and spectral consistency across domains. Focused on stabilizing adversarial training and faithful spectral reconstruction for remote-sensing data.

Collaborated on a mental wellness companion application using MERN stack, recognized as a top project by Google Developers Club Delhi. Designed responsive UI with React and TailwindCSS, integrated Firebase analytics and MongoDB for secure data management. Prepared stress analysis techniques using sentiment analysis and mouse tracking with 92.71% accuracy.

Built a real-time, cross-platform AI search assistant with Flutter and FastAPI. Created a custom ranking algorithm using sentence transformers achieving 85% accuracy in source ranking, with WebSocket streaming for live responses and a RAG pipeline combining the Gemini API with Tavily Search for accurate information synthesis.

Innovated a raw-to-dense feature pipeline, transforming raw bid data using contrastive embeddings & autoencoders, followed by an ANN-based feature compression for efficient representation. Devised a low-latency RTB model predicting bid price and bidding decisions within 5ms per request, optimized via Grid Search & Adaptive Learning Rate Scheduling. Achieved 82% classification accuracy, scaling to handle 100K+ bid requests per second in large-scale ad exchanges.

Prepared an ML-based system to identify distinct market states from high-frequency financial data. Engineered 20+ custom features from order book data, consolidated 3 clustering algorithms achieving 87% silhouette score. System successfully identified 4 distinct market regimes with 92% classification accuracy, reducing strategy drawdowns by 15% through adaptive position sizing. Analyzed 500,000+ data points across multiple timeframes to enable real-time regime classification within 50ms.