Economics (BS-MS) at IIT Kanpur, final year. Minor in Machine Learning Algorithms from CSE, IITK. I build AI agents and their workflows. I also build econometric and machine learning models and the software around them.
Both halves come from the same habit. I want to know where a model stops working. So I price the same option three ways and see where the numbers split. I nowcast GDP from data that arrives late and incomplete. I build agents that are allowed to say no instead of guessing.
I try to report what actually happened. That includes the parts that did not work.
Graduating 2027. Looking for roles in quantitative research, econometrics, data science and AI engineering.
Riva and Riva Snap A companion app for people on GLP-1 medication. It ships with the medicine. Patients track weight, doses, food, water and how they feel each week. The iPhone app is SwiftUI. The backend takes a photo of a meal and returns the dish, the portion and the nutrients. It is grounded in USDA food data instead of guessed.
Prep OS An offline first Flutter app I use every day. It records a lecture in the background and transcribes it. It reads photos of the board with a vision model. Then it chains LLM stages into proper revision notes. It handles token budgets, rate limits across providers and partial results when something fails. I built it because I needed it.
Personalised Offer Optimisation Decides which cashback offer a payment platform sends each user each week, learned from logged data alone. The policy beats the shipped heuristic by Rs 2.91 per decision while spending 37% less on coupons. A GAN world model and a PPO agent came first and both lost to a plain regression. The diagnostic that explains why is the part worth reading.
Nowcasting Indian GDP with a Dynamic Factor Model Estimates this quarter's GDP before the official number is out. It works from monthly indicators that arrive at different times and are often incomplete. This is the problem central banks actually deal with.
MuleGuard Finds money mule bank accounts using a tabular model and a graph neural network together. It works across banks without sharing raw data. Fraud is a graph problem before it is a classification problem.
European option pricing, three ways Black Scholes, binomial trees and Monte Carlo written side by side in MATLAB. Where the three prices disagree is where the assumptions stop holding.
ARIMA-LSTM for crop prices Classical time series and a neural model combined, with random forest feature engineering. The price series are seasonal and sparse and full of shocks.
AI agents and workflows LLM pipelines. Multi stage orchestration. Tool use and guardrails. Prompt design. Speech to text. Vision models. RAG. Streaming, retries and rate limits.
Modelling Econometrics. Time series. Dynamic factor models. ARIMA. Causal inference. Off policy evaluation. Reinforcement learning. Option pricing. Panel data.
Machine learning PyTorch. Graph neural networks. scikit-learn. Neural forecasting. Feature engineering.
Languages and tools Python. TypeScript. MATLAB. R. SQL. Swift. Dart and Flutter. FastAPI. Next.js.
LinkedIn: ashutoshanandmehta
Email: anandmehtaashutosh@gmail.com