Neel Mishra

Career

Experience

3+ years building production ML systems, GenAI products, and data-intensive pipelines across Microsoft, Myntra, and Tekion. 2 peer-reviewed publications at IJCNN (IEEE).

Timeline

Applied Scientist II · Microsoft

Sep 2025 - Present · Bangalore

  • Building heterogeneous knowledge graph representations to propagate verdicts across multiple indicators (apps, IPs, emails, sessions), improving threat attribution accuracy.
  • Driving the design and implementation of graph-based threat detection and explainability systems using GNN architectures (GraphSAGE, GCN), risk propagation techniques, and embedding-based models.
  • Leading the development of scalable, low-latency Azure Synapse data pipelines to productionize machine learning models for real-time threat detection and response.
  • Knowledge Graphs
  • GNN
  • Azure Synapse
  • Threat Detection

Data Scientist · Myntra

Sep 2024 - Jul 2025 · Bangalore

  • Created and deployed the Maya GenAI chatbot, driving a 58.86% increase in engaged MAU and a 31.74% improvement in repeat usage through a multi-agent architecture with dedicated personalization and search agents.
  • Enhanced Myntra’s recommendation system, achieving 0.81 nDCG and 0.61 MAP@15 for personalized search by fine-tuning the CLIP model on 1.4M image–description pairs.
  • Built scalable PySpark pipelines to train Word2Vec embeddings on 500M+ records, leveraging custom evaluation metrics to reduce embedding dimensionality from 100 to 22 while preserving semantic quality.
  • GenAI
  • Multi-Agent
  • CLIP
  • PySpark
  • Word2Vec

Data Scientist · Tekion

Jul 2023 - Aug 2024 · Bangalore

  • Fine-tuned LLM models tinyLlama, Llama2, and Zephyr to enhance the email generation process, improving user experience for multiple automotive dealerships.
  • Achieved a 98.08% reduction in execution time for fine-tuned LLMs through advanced techniques like paged attention and post-training quantization.
  • Conducted advanced statistical analysis to validate the improved performance of LLMs, achieving 21% higher similarity and 35% increased relevance in generated emails.
  • Developed a robust RAG question-answering chatbot for vehicular marketing, processing comprehensive data for over 800 car models using Langchain and advanced features.
  • LLM Fine-tuning
  • RAG
  • Langchain
  • Quantization

Research · IIIT Hyderabad

2019 - 2023

  • Collaborated with Principal Applied Scientists at Microsoft India to develop a minimax optimizer for GANs, significantly improving inception scores while reducing execution time to 0.184 seconds.
  • Implemented GPU-accelerated PyTorch solvers for Conjugate and Bi-Conjugate Gradient methods, achieving over 96% runtime reduction compared to standard NumPy implementations.
  • Authored the paper “Angle based learning rate for gradient descent,” presented at the International Joint Conference on Neural Networks (IJCNN 2023, CORE A), proposing a novel learning rate adaptation method.
  • Authored the paper “A Gauss-Newton Approach for Min-Max Optimization in GANs” under the Microsoft Academic Partnership Grant (MAPG), improving GAN optimization efficiency using the Gauss-Newton method (WCCI 2024).
  • GANs
  • PyTorch
  • IJCNN
  • WCCI
  • Research

Publications

Google Scholar →
WCCI 2024 · IEEE

A Gauss-Newton Approach for Min-Max Optimization in Generative Adversarial Networks

N Mishra, B Mishra, P Jawanpuria, P Kumar

A novel first-order method for training GANs using a modified Gauss-Newton method to approximate the min-max Hessian. Achieves highest inception score for CIFAR10 among all compared methods including state-of-the-art second-order methods.

arXiv IEEE CORE A
IJCNN 2023 · IEEE

Angle Based Dynamic Learning Rate for Gradient Descent

N Mishra, P Kumar

A novel adaptive learning rate using the angle between current and orthogonal gradients. Outperforms Adam, AdaGrad, and RMSProp on CIFAR10/100 across ResNet, DenseNet, EfficientNet, and VGG architectures.

arXiv IEEE CORE A

Impact Snapshot

58.86%

Engaged MAU uplift for a GenAI chatbot

Myntra

0.8s

LLM inference time after optimization

Tekion

500M+

User events processed via PySpark

Myntra

0.184s

Minimax GAN optimizer runtime

IIIT Hyderabad