Mohd Maaz Khan

Education

Experience

• Research Intern

  Oct 2025 – present

  Department of Chemical Engineering, IIT (ISM) Dhanbad · with Prof. Saurav Sengupta

  • Reconstructed the Lorenz system with two sparse regression techniques (SINDy vs. ISINDy) and compared them.
  • Built a surrogate model for 1D diffusion using DeepONet.
  • Modelled a spring-mass system with Hamiltonian Neural Networks and compared them with baseline NNs on accuracy and physical consistency.

Projects

• Enterprise RAG System with LLM-as-a-Judge

  In progress

  Hybrid-retrieval RAG with an LLM judge that scores every answer for faithfulness, relevance and toxicity.

  • Hybrid retrieval: dense embeddings (text-embedding-3-small) in Qdrant, combined with BM25 keyword search.
  • LLM-as-a-Judge: a second model with Pydantic structured outputs scores each answer for faithfulness, context relevance and toxicity.
  • Observability: async PostgreSQL logging of query trails, latency and scores.
  • Frontend: React + TypeScript with streaming answers, citations and Recharts analytics.
  • Shipping: Docker Compose, with CI/CD on GitHub Actions.

  Stack

  Python · FastAPI · React · TypeScript · LangChain · PostgreSQL · Qdrant · Docker

  Code

  github.com/maaz7409/SmartRAG

• E-Commerce Customer Churn Prediction

  Live

  Stacking ensemble that predicts which e-commerce customers will churn, deployed with a live app.

  • 96%Accuracy
  • 0.93Macro F1
  • Preprocessing: KNN imputation, winsorization, one-hot encoding, and SMOTE for class imbalance.
  • Model: a stacking ensemble (KNN, SVM, XGBoost, Random Forest) with a Logistic Regression meta-model.
  • Results: 96% accuracy and 0.93 macro F1.
  • Deployment: FastAPI backend and React frontend, both on Vercel.

  Stack

  Python · Scikit-Learn · Imbalanced-Learn · XGBoost · FastAPI · React · Docker

  Code

  github.com/maaz7409/customer-churn-predictor

  Live

  ecommerce-customer-churn-predictor.vercel.app

• Automated Email Classification Engine

  SVM spam/ham classifier with custom parsing and zero train/test leakage.

  • 0.97Macro F1
  • 97%Accuracy
  • Custom email parsing and cleaning.
  • Vocabulary building and count vectorization.
  • Linear SVM classifier: 0.97 macro F1 and 97% accuracy, with no train/test leakage.

  Stack

  Python · Scikit-Learn · SVM · NLP

  Code

  github.com/maaz7409/email_classifier

• EHR Patient Readmission

  In progress

  Predicting 30-day hospital readmission from electronic health records.

  Stack

  Python · Scikit-Learn

• Amazon ML Challenge 2026

  Team solution for business entity resolution, ranked 287 of ~25,000 teams.

  • 0.988Macro F0.5 (public LB)
  • 287 / ~25,000Rank
  • Three blockers (word TF-IDF, character TF-IDF, a fine-tuned multilingual bi-encoder) plus a learned pruner keep ~11.5 candidates per record at 99.5% recall.
  • A LightGBM model over 119 similarity features and three fine-tuned mmBERT cross-encoders score the candidates.
  • A second, graph-aware LightGBM combines their scores; one global rule tuned on validation F0.5 emits the matches.
  • Two self-training rounds adapt the cross-encoders to France, which has no labels.

  Stack

  Python · PyTorch · Transformers · Sentence-Transformers · LightGBM · Scikit-Learn · Pandas

  Code

  github.com/maaz7409/AmazonMLChallenge-2026

  Live

  ballin-amazonml2026.vercel.app

Research

• Inverse Heat Transfer with PINNs

  Recovering unknown heat flux and material properties with a physics-informed neural network.

  • 99%Accuracy (time-invariant flux)
  • 97%Accuracy (time-variant flux)
  • ~10× fasterConvergence vs. paper
  • Synthetic training data generated with Green's functions.
  • A PINN with a custom learning-rate scheduler that co-optimises the network weights and the thermophysical parameters.
  • 99% accuracy for time-invariant flux and 97% for time-variant flux.
  • Converges ~10× faster than the original paper.
  • The repo has the data generation (GT.ipynb), the training (iht_with_pinn.ipynb) and a full report (report.pdf).

  Stack

  Python · PyTorch · SciPy

  Code

  github.com/maaz7409/Inverse_Heat_Transfer_with_PINNs

• CFD with Python

  A fluid-flow solver written from scratch that shows Von Kármán vortex shedding.

  • 0.159Strouhal number at Re = 200
  • It solves the 2D incompressible Navier–Stokes equations at $Re = 200$, in a channel $20D$ long and $10D$ wide with the cylinder at one fifth of the length.
  • Chorin's projection method on a staggered grid, with the cylinder added by an immersed boundary method.
  • Numba compiles the grid loops to parallel machine code.
  • The run saves velocity and vorticity frames, a video of the simulation, and a dataset (.npz) for later surrogate modelling with a PINN.
  • The repo has the explanation (Explanation.ipynb) and the code (CFD.ipynb) side by side.

  Stack

  Python · NumPy · SciPy · Numba · OpenCV · Matplotlib

  Code

  github.com/maaz7409/CFD-with-Python

• Hamiltonian Neural Networks for a Spring-Mass System

  A network that learns the Hamiltonian, compared with a baseline NN on accuracy and energy conservation.

  • 50 points over one period: inputs $(q, p)$, targets $(\dot{q}, \dot{p})$.
  • Noise to test robustness: 2.5% of the standard deviation added to $(q, p)$ and 5% to $(\dot{q}, \dot{p})$.

  Stack

  Python · PyTorch · NumPy · Matplotlib

  Code

  github.com/maaz7409/HNN

• DeepONet Surrogate for 1D Diffusion

  Learning the operator that maps a diffusion coefficient D(x) to the full solution u(x, t).

  • 2.56%Relative L2 error (Neumann)
  • 3.98%Relative L2 error (Dirichlet)
  • Dirichlet (open ends): $u = 0$ at $x = 0$ and $x = 1$, so the concentration drains away.
  • Neumann (closed ends): $\partial u/\partial x = 0$ at both ends, so nothing leaves and the profile flattens towards its mean.

  Stack

  Python · PyTorch · Scikit-Learn · NumPy · Matplotlib

  Code

  github.com/maaz7409/DeepONet

• SINDy vs. ISINDy on the Lorenz System

  Recovering the Lorenz equations from data with sparse regression, and making it robust to noise.

  • 98.5%Coefficient error at 5% noise: SINDy
  • 3.3%Coefficient error at 5% noise: ISINDy

  Stack

  Python · NumPy · SciPy · Scikit-Learn · Matplotlib

  Code

  github.com/maaz7409/SINDy_vs_ISINDy

Open source

50 merged pull requests to 11 open-source projects.

Every pull request: maaza-codes.vercel.app/contributions

Competitive programming

Details: maaza-codes.vercel.app/CP.md

Skills

Achievements

• American Express Strategy Track Challenge

  Top 10% of 17,000+ participants. Built a customer-profitability ranking model with an accuracy of 0.882.

• Amazon ML Challenge 2026

  Team result: public leaderboard macro F0.5 0.988, rank 287 of ~25,000 registered teams (business entity resolution). See projects/amazon_ml_challenge.py.

• AMS DERIVE (PRIOR)

  Quantitative problem solving: financial modelling, Bayesian statistics, stochastic modelling and probability.

Interests

• chess
• cricket
• table tennis