Biography & Foundations

About Me

👋 Hi there! I'm Sushovan Saha — currently working as a founding AI Engineer at a Stealth Startup, building production-grade intelligent systems with deep focus on Large Language Models (LLMs), Generative AI, Voice Agents, Deep Learning, Natural Language Processing (NLP), and Computer Vision (CV).

🎓 I hold an M.Tech in Data Science from IIT Guwahati and I'm currently a Kaggle Notebook Expert with a strong inclination toward solving real-world challenges using intelligent systems.


Technical Competencies

Engineering Capabilities

  • LLMs & Generative AI: Proficient in RAG, GraphRAG, Agentic RAG, MCP, LangChain, LangGraph, Google ADK, and building Agentic Workflows
  • Voice Agents & Real-Time AI: Building end-to-end real-time Voice Agents using LiveKit, WebRTC, low-latency audio pipelines, and tool-augmented conversational agents
  • Model Fine-Tuning & Post-Training (Reasoning & RL): Hands-on experience in fine-tuning models on custom datasets with SFT (Supervised Fine-Tuning) and GRPO (Group Relative Policy Optimization) on Qwen3.5 4B, leveraging HuggingFace, Unsloth, PyTorch, and PyTorch Lightning
  • Applied ML & Deep Learning: End-to-end project experience across NLP, Computer Vision (CV), and Time Series domains
  • Time Series & Forecasting: Statistical and deep learning models for trend analysis, predictive modeling, and demand Forecasting
  • MLOps & Scalable ML Systems: Focus on production-grade pipelines, model deployment, and monitoring with MLflow
  • API & Deployment: Experience in deploying models as APIs using FastAPI and deploying on cloud platforms like AWS and GCP
  • CI/CD: Experience in setting up CI/CD pipelines using GitHub Actions
  • Data Science: Experience in data analysis, statistical analysis, data visualization, and data preprocessing using Pandas, NumPy, Matplotlib, Seaborn, and Scikit-learn

Tooling & Platform Stack

Frameworks & Ecosystem

LiveKit LangChain LangGraph CrewAI Hugging Face PyTorch Lightning Keras Kaggle Google Colab Unsloth

Beyond the Code

Personal Pursuits

  • Contributing to the ML community on Kaggle
  • Writing practical ML blogs on Medium
  • Exploring new AI tools, frameworks, and academic papers to stay at the cutting edge
Quantitative Intelligence • Indian Equities

AI Stock Research Agent

High-probability range breakouts, financial health diagnostics, and deep fundamental screening across NIFTY500 and MICROCAP250.

Stock Universe:
Click 'Run Scan' to scan stocks breaking out with high volume.

Harry Potter & Ancient Mythology

What secrets of the wizarding world shall we uncover?

Cross-examine Harry Potter lore, spells, Horcruxes, and Indian ancient epics via Pinecone vector retrieval

Airbnb MCP • Weather Real-Time Pipeline

Autonomous Tour Planner

What’s on the agenda today?

Query live Airbnb accommodations via MCP and real-time weather forecasts to curate verified vacation itineraries

Low-Latency WebRTC • Silero VAD • Groq

Real-Time Voice Agent

Ultra-low-latency conversational intelligence with full-duplex audio streaming and dynamic tool dispatch.

Disconnected
Room: Ready
Pipeline: Silero VAD + Groq + Cartesia
Tap Call to Start
Real-time WebRTC Voice AI
Executing Tool
Computer Vision • Flickr27 Transfer Learning

YOLO Object Detection

State-of-the-art brand mark localization and real-time bounding box prediction on custom imagery.

🚀 YOLOv8.1: The Latest in Object Detection

  • 🆕 YOLOv8.1 is out! The newest update in the YOLO series, maintaining its position as state-of-the-art for Object Detection, Instance Segmentation, and Classification.

⚠️ Main Challenge: Custom Dataset Preparation

  • 🔍 Dataset Selection: Using Flickr27 as our image dataset.
  • 📸 Flickr27 Overview: Contains 27 different brand logos, perfect for training YOLO on custom data.
  • 💼 Custom Dataset Prep: The most crucial step in training YOLO models.
  • 🛠️ Get Ready to Train: With YOLOv8.1 and Flickr27, you'll be well-equipped to handle custom object detection tasks!
📁
Upload an Image containing Brand Logo
Limit 200MB per file • JPG, JPEG, PNG
Deep Learning Benchmarks • Transfer Learning

Image Classification Matrix

Comparing inference latency, parameter footprint, and top-1 accuracy across modern vision architectures.

🚀 Exploring Image Classification Models with Transfer Learning!

  • 🆕 Models Evaluated: We've assessed the following state-of-the-art models using the Flickr27 dataset:
    • Xception
    • InceptionV3 🌈
    • MobileNetV2 📱
    • EfficientNet

🔍 Performance Aspects Compared

  • ⚡ Inference Time: How quickly each model makes predictions.
  • 📦 Model Size: The storage requirements of each model.
  • 🔢 Number of Parameters: The complexity and capacity of each model.
  • 📈 Accuracy: How well each model performs in classifying images.
  • 🏷️ Predicted Class: The class each model predicts for the input images.
  • 📚 Understanding Trade-Offs: These comparisons will help us understand the trade-offs between speed, efficiency, and accuracy for each model.

💡 You can train your own model on any dataset following the link below:
Train Your Model

🚀 To see How it works..

📁
Upload an Image containing Brand Logo
Limit 200MB per file • JPG, JPEG, PNG
Unsupervised Machine Learning • Density vs Centroids

Clustering Playground

Explore dataset partitioning via K-Means and density-based spatial clustering (DBSCAN).

💡 Why don't you try DBSCAN..!!
Open-Source Community & Research

Connect & Collaborate

Connect with me across platforms, explore open-source code repositories, or read my technical publications.

Connect with me across platforms, view my open-source code repositories, or read my technical AI/ML publications:


GitHub Stats

Username: ambideXtrous9
Name: Sushovan Saha
Loading public repos...
Loading followers...
Identity • Role-Based Access Control

User Authentication

PostgreSQL JWT Bearer Authentication and persistent session checkpointing.

🛡️ PostgreSQL & JWT Identity System
Role-Based Access Control and PostgreSQL session checkpointing. Visitors have instant guest access enabled.
RBAC Enforced

Sign In to your Account