Image Captioner — Misa
Three AI projects in one: a DeepSeek-powered chatbot, an unbeatable Tic-Tac-Toe AI, and a BLIP-based image caption generator.
Built by Sandeep Prajapati
PythonBLIPHugging FaceDeepSeek R1JavaScriptMinimax
← Projects→→→
Overview
Three AI mini-projects bundled into one, built during the MISA program — each exploring a different corner of applied AI: a conversational chatbot, a game-playing agent, and a vision-language captioner.
What I Built
1. DeepSeek-powered chatbot — a conversational assistant built on the DeepSeek R1 model.
2. Unbeatable Tic-Tac-Toe AI — a game agent using the MinimaxA game-tree search for two-player games: assume the opponent always plays their best move, and pick the move whose worst case is best for you. algorithm; it never loses.
3. BLIPBootstrapping Language-Image Pre-training — a family of vision-language models used for tasks such as image captioning and visual question answering. image caption generator — a vision-language model (BLIP via Hugging Face) that generates natural-language descriptions of uploaded images.
Tech Stack
AI / ML: BLIP, Hugging Face Transformers, DeepSeek R1
Algorithms: Minimax (game AI)
Languages: Python, JavaScript
Why It Matters
A compact tour of three AI paradigms — LLMLarge language model — a neural network trained on large amounts of text to predict the next token, used for generation, summarisation, classification and more. chat, classical game-search, and multimodalAble to work with more than one kind of input — such as text plus images or audio — in the same model. vision-language — implemented end to end.
Related projects
Tic-Tac-Toe AI
Web-based Tic-Tac-Toe with an unbeatable AI opponent powered by the Minimax Algorithm with Alpha-Beta Pruning.
Noloop
A glass-box medical claims adjudication and fraud-defense platform — an AI agent pipeline turns unstructured claim documents into explainable, role-specific decisions for patients, hospitals, and insurers.
RAG Document Assistant
A production-style Retrieval-Augmented Generation system — documents are chunked, embedded with SentenceTransformers, indexed in Endee's HNSW vector database, and retrieved via cosine-similarity k-NN search to ground Groq's Llama 3.3 70B in a user-provided knowledge base with full source attribution.