Building data engines & agentic intelligence.
I'm Soumotanu Mazumdar. I build at the intersection of modern cloud data warehousing (Snowflake) and autonomous generative AI systems (Google Gemini & Vertex AI).
My work spans architecting modular data pipelines, developing autonomous reasoning agents with grounded memory, and designing high-performance machine learning and computer vision systems.
// Selected Works
Code, Architecture & Implementations
An intelligent autonomous concierge built on the Google Agent Development Kit (ADK) and deployed to Vertex AI Reasoning Engine. Combines persistent state memory banking, custom sandbox code analytics, real-time weather retrieval, and Vertex AI RAG Corpus grounding in botanical literature.
Real-Time Object Tracking with YOLOv5 & DeepSORT
High-throughput tracking pipeline combining YOLOv5m object detection with DeepSORT and OSNet ReID for persistent trajectory tracking across heavy crowd occlusions on the MOT20 benchmark. Features 8D Kalman state estimation and cascaded cosine matching.
Production-grade ELT analytics platform implemented on Snowflake Cloud Data Warehouse using dbt Core. Designed following Kimball dimensional modeling paradigms with multi-layer DAG transformations (staging, intermediate, dimension/fact marts), Slowly Changing Dimensions (SCD Type 2), Jinja macros, and automated schema testing.
Sign Language & 3D Hand Pose Detection
Real-time spatial geometry pipeline using MediaPipe 21 3D joint landmark extraction and OpenCV. Computes relational euclidean vectors across fingers to classify sign language gestures in real time using custom ANN and CNN models.
Transformer Chatbot from Scratch in PyTorch
Mathematical first-principles implementation of the vanilla Transformer architecture in PyTorch. Includes 8-head scaled dot-product attention, sinusoidal positional embeddings, and Noam learning rate scheduling on dialogue datasets.
Neuroevolution Flappy Bird AI
Genetic Algorithm implementation in Pygame that evolves neural network weights and biases across successive bird populations. Includes real-time rendering toggles and headless unlimited FPS mode for rapid evolutionary convergence.
Interactive Digit Recognizer Canvas
Interactive drawing surface built in Pygame paired with a 99.65% test accuracy Convolutional Neural Network trained on MNIST. Converts user-drawn raster sketches to normalized 28×28 tensor matrices for sub-millisecond inference.
// Approach & Craft
How I Approach Engineering
Data Platforms as Code
In Snowflake, architecture is about predictable cost and strict lineage. I design dimensional data marts with idempotent transformations, automated schema validation, and zero-copy cloning for risk-free sandbox testing.
Reasoning Over Prompting
LLMs are stochastic; enterprise systems cannot be. I architect multi-agent systems with explicit state machines, persistent memory banks, external tool grounding (RAG), and deterministic fallbacks rather than relying on brittle prompts.
Cross-Domain Breadth
Understanding computer vision matrices, RTMP media protocols, and Godot game loops prevents engineering tunnel vision. Solving problems in graphics and systems directly informs better database performance and AI tool design.
// Tooling Matrix
Technologies I Use
- Snowflake Cloud
- dbt (Data Build Tool)
- PostgreSQL
- Google Cloud (GCP)
- Firestore
- Google Gemini API
- Vertex AI Reasoning
- Vertex RAG Engine
- MediaPipe & OpenCV
- TensorFlow / PyTorch
- Python (3.10+)
- SQL (Snowflake / ANSI)
- GDScript (Godot)
- Bash / Shell
- C++
- GitHub Actions CI/CD
- Docker & Linux
- FastAPI Microservices
- Data Quality & Testing
- Git Workflows
Let's build something enduring.
Whether you want to discuss Snowflake warehouse optimizations, multi-agent RAG implementations, or open-source software, my inbox is open.