Amit Kamble —Building LLM, CV, and Agentic Systems.

A scientifically grounded, multi-metric RAG application built for the Darukaa.Earth Biodiversity Intelligence Challenge with RRF hybrid retrieval and schema-enforced causal reasoning.

• Deterministic Memory & Clarification Engine: Discrete Pydantic state tracking across multi-turn sessions with automatic intent-gating to halt LLM and ask follow-up questions if context is insufficient
• Hybrid RAG with RRF: Combines dense vector embeddings (LanceDB) and full-text lexical search (BM25) over 2,961 indexed FAO/IPCC document chunks
• Structured Causal Reasoning & Resilient Fallbacks: Rigid JSON output schema with directional impact metrics (📈/📉), exact page citations, and automatic multi-tier Gemini model fallback chain

A civic operating system for sanitation reporting, route optimization, and epidemic prevention.

• Photo + voice reporting for low-literacy access
• 95%+ spam detection accuracy
• 30% fuel savings and 40% labor efficiency improvement

A deep learning pipeline that detects C/C++ vulnerabilities with CodeBERT + LoRA and AST scoring.

• Best F1 score: 74.32% on held-out test set
• 87.94% recall with tuned threshold
• Covers 12+ CWE categories with AST ensemble scoring

Multimodal medical AI research combining radiology text priors with MRI features for tumor segmentation.

• Vision + language fusion pipeline
• BraTS 2020 and TextBraTS datasets
• Patient-level split with zero data leakage