Aditya Batlawala.

▸ 87% top-1 accuracy across 101 food classes (~101k images)

Snap a photo of any meal — or point your webcam at it — and the app tells you what it is, how confident it is, and instantly pulls up calories, protein, carbs and a full recipe. Recognizes 101 different dishes. Under the hood: a fine-tuned ConvNeXt-Tiny vision model trained in a production-style five-stage pipeline with modern augmentation and evaluation, deployed live with Docker.

▸ R² = 0.81 on held-out test data

Enter a Berlin apartment's size, rooms, heating type and zip code — get an instant price estimate, trained on real ImmoScout24 listings. Useful for anyone wondering if a listing is over- or under-priced. Under the hood: an end-to-end ML system that trains four model families, automatically picks the best one, tracks every experiment in MLflow, and ships to AWS with CI/CD.

Paste a Flipkart product link and instantly see what hundreds of customers actually think — overall sentiment, positive vs. negative breakdown, and interactive charts — without reading a single review yourself. Under the hood: live web scraping with BeautifulSoup and VADER sentiment analysis served through a Flask app.

Building an AI assistant that answers questions from your own documents — accurately, with sources — instead of hallucinating. Think: ask anything about a 200-page PDF and get a grounded answer in seconds. Under the hood: Retrieval-Augmented Generation with LangChain, comparing open-source LLMs (Llama via Hugging Face and Groq), embedding models and vector databases, with quality measured by faithfulness and retrieval metrics.