17+ models across 4 providers. Benchmark mode fires the same prompt at all models simultaneously with live result cards, latency bars, and winner badges (⚡ Fastest · 💰 Cheapest · 📝 Most Detailed). Private mode routes sensitive data to local Ollama only — never leaves your machine. Reasoning chains from Qwen3 <think> blocks extracted and shown as collapsible toggles.

Agents are assigned distinct roles and collaborate through a shared task planning loop. Each agent uses tool-use to fetch live data, reason over it, and pass structured output to the next stage — resulting in a fully autonomous research-to-report pipeline.

Resolved real-world environment compatibility challenges across bitsandbytes, tokenizers, and transformers versioning. Training tracked with Weights & Biases. The fine-tuned adapter and base model are published to HuggingFace Hub with a live Gradio inference demo.

Solved key production challenges including async event loop conflicts in FastAPI, streaming SSE via LangGraph ReAct, and ChromaDB/BM25/FlashRank hybrid retrieval tuning. Powered by Groq's llama-3.1-8b-instant with HuggingFace embeddings.

Structured as a reproducible, production-mimicking workflow with modular stages for data ingestion, training, evaluation, and monitoring. Evidently AI generates drift reports automatically on new data batches, with all runs logged and versioned in MLflow.

End-to-end diagnostic pipeline built for early TB screening. Custom-trained TBNet CNN architecture analyses chest X-rays, highlights infected regions via Grad-CAM heatmaps, and generates professional PDF diagnostic reports. FastAPI backend with optimised inference pipeline and an intuitive web interface.

AI tutor capable of analysing images, PDFs, and text files using Llama 3, Qwen-VL (vision), and Groq models. Includes chat history export, file previews, secure secret handling, and custom UI components.

PTE Academic mock system with Aptitude, Listening, and Reading modules. Features automated question generation via Groq & Gemini, strict time tracking, dynamic content validation, analytics-driven scoring, and a premium custom UI. Secure API handling and modular code architecture.

Trained on 40,000+ news articles. Features real-time predictions, analytics dashboard, prediction history, and dual-model verification using both a custom neural network and Gemini AI.

Trains 7 classifiers on real IoT sensor data. MobileNetV2 fine-tuned via transfer learning achieves 96.98% validation accuracy. YOLOv8 draws real-time bounding boxes on uploaded images. Full-stack Django app with role-based user management, live ML training, sensor prediction, and CNN+YOLO inference — deployed on Railway.