Mixtral 8x7B — A Deep Dive
A detailed comparison of Mixtral 8x7B with LLaMA 2, and an implementation of an optimized Mixture of Experts (MoE) layer in PyTorch.
Read on SubstackMachine Learning Engineer
Specializing in Large Language Models and RAG systems. Building production ML infrastructure and sharing deep technical implementations from first principles.
About Me
I'm a Machine Learning Engineer specializing in Large Language Models and Retrieval-Augmented Generation (RAG) systems. Currently at Verloop.io, I architect end-to-end RAG pipelines that power customer support for enterprise clients across e-commerce, banking, and healthcare sectors.
I believe in understanding AI systems at a fundamental level—from transformer architectures to retrieval optimization. Through my technical blog, I break down complex topics like Rotary Positional Embeddings (RoPE), attention mechanisms (MHA, GQA, MLA), and modern LLM architectures, providing complete PyTorch implementations from scratch.
My work bridges cutting-edge research and production systems. I focus on building with clarity, sharing knowledge openly, and making advanced ML techniques accessible to practitioners. Previously at Monsoon CreditTech, I built credit risk models achieving 40% reduction in delinquency rates.
Skills & Expertise
Technologies and areas I work with
LLM & AI
- Large Language Models
- RAG Systems
- Retrieval Systems
- Vector Databases
- Embedding Models
- Prompt Engineering
- Transformer Architecture
ML Frameworks
- PyTorch
- Hugging Face Transformers
- Scikit-learn
- XGBoost
- Weights & Biases
- SHAP
- Optuna
Infrastructure
- Docker
- Kubernetes
- gRPC
- Kafka
- RabbitMQ
- Redis
- PostgreSQL
- MySQL
Languages & Web Frameworks
- Python
- SQL
- FastAPI
- Django
- Gradio
Cloud & DevOps
- Google Cloud Platform
- CI/CD
- GitHub Actions
- Git
Expertise
- Deep Learning
- Natural Language Processing
- Mathematics
- Microservices Architecture
Experience
My professional journey and key achievements
SDE - Machine Learning
Oct 2024 - PresentVerloop.io
Building production RAG systems and LLM infrastructure for enterprise customer support automation.
- Architected end-to-end RAG pipeline supporting multiple document formats (PDF, Markdown, Text) for enterprise clients
- Improved retrieval accuracy by 15% (Recall) and 5% (MRR) through systematic evaluation of embedding models and rerankers
- Built production microservices with Docker, Kubernetes, gRPC, and integrated Weaviate vector database
- Optimized LLM context windows, reducing latency and API costs while maintaining answer quality
Machine Learning Engineer
Feb 2022 - Aug 2024Monsoon CreditTech
Developed ML models for credit risk assessment and fraud detection in fintech.
- Reduced loan delinquency rates by 40% through credit risk model using XGBoost and Bayesian optimization
- Implemented comprehensive MLOps monitoring suite with drift detection (PSI, KS test) and SHAP explainability
- Deployed scalable REST APIs using FastAPI and Django, containerized with Docker on GCP Cloud Run
- Built fraud detection model using deep isolation forest, identifying 30% of anomalies in top risk decile
Bachelor of Science in Computer Science
Jul 2016 - Aug 2020BML Munjal University
Major in Computer Science and Engineering. Published research in Nature Scientific Reports.
- Published: "Machine learning predicts live-birth occurrence before IVF treatment" in Nature Scientific Reports (2020)
- Focused on Machine Learning, Deep Learning, and Software Engineering
Featured Blogs
Latest technical articles combining theory and practice
Building LLaMA Model: A Deep Dive
Step-by-step guide to building the LLaMA model from scratch in PyTorch, with in-depth explanations of each essential component.
Read on SubstackAttention Mechanisms & KV Cache: A Deep Dive
Understanding the evolution from Multi-Head Attention to modern inference optimizations
Read on SubstackGet In Touch
I'm always interested in discussing LLMs, RAG systems, ML engineering, or potential collaborations. Feel free to reach out!