Pawan Singh
Ahirwar
I build and ship machine-learning systems end to end — from CNN image classifiers on 800K+ images in production at the Ministry of Agriculture, to self-correcting RAG pipelines with reranking and metadata filtering. M.Tech from IIT Bombay.
Who I am
I'm an AI/ML Engineer with an M.Tech from IIT Bombay and a background in Mechanical Engineering — a path that taught me to break complex, messy problems down into systems that actually work.
Today I work at the Ministry of Agriculture, Government of India, building computer-vision and deep-learning models that run at national scale across hundreds of thousands of crop images. I care about taking ambiguous goals and large, imperfect datasets and turning them into things that ship and create real impact — especially in agriculture and underserved communities.
Lately I've gone deep into generative AI, building retrieval-augmented systems that are precise, grounded in their sources, and efficient enough to run on modest hardware. I like building the whole pipeline, end to end.
Projects & systems I've built
Digital Crop Survey — CNN Image Classifier
Gov. of IndiaProduction computer-vision model classifying 8+ crop types for the Ministry of Agriculture's nationwide crop survey, with preprocessing pipelines built to handle very large agricultural image datasets for government deployment.
Advanced RAG — LangGraph + Qdrant + BGE Reranker
Open sourceA self-correcting Retrieval-Augmented Generation system built as a stateful graph (retrieve → rerank → grade → generate). Cross-encoder reranking, Qdrant metadata filtering, page-level citations, and a "not found" guard so it won't hallucinate. Runs fully offline.
Local RAG Chatbot — Chat with your PDFs
Open sourceA fully local RAG chatbot that answers questions grounded in an uploaded PDF — embeddings, vector search, and a local LLM, all running offline with no API keys. The foundation I later rebuilt with reranking and a graph pipeline.
Cyberbullying Recognizer & Summarizer
NLP · 89% accAn NLP classifier for cyberbullying detection on social-media text — ETL and EDA with SpaCy, feature extraction, a Decision Tree model at 89% accuracy, deployed as a Flask Web API and containerized with Docker.
What I work with
Computer Vision & Deep Learning
NLP
Generative AI
Machine Learning
Data & Deployment
Cloud & Infrastructure
Where I've worked
- Digital Crop Survey: CNN crop classifier (92% accuracy, 800K+ images) with large-scale preprocessing for government deployment.
- National Pest Surveillance System: multi-class deep-learning model for pest & disease detection at 87% test accuracy.
- Advanced RAG tooling for natural-language querying of agricultural reports.
- Thesis presented at the 5th International Conference on Waste Management, IIT Guwahati.
Let's build something
Open to AI/ML, computer vision, and applied GenAI roles. The fastest way to reach me is email or LinkedIn.