Rajarshi Das

Rajarshi Das

AI/ML Engineer specializing in the local open-weights LLM ecosystem, agentic AI workflows, and production-grade machine learning systems.

View Projects ⚡ Launch Void Drop P2P 🛡️ Launch PrivaSizer View Resume

About Me

I am an AI/ML Engineer specializing in the local open-weights LLM ecosystem, agentic AI workflows, and production-grade machine learning systems. I have significant experience in running, benchmarking, and optimizing open-source models using Ollama, llama.cpp, and LM Studio. My background covers model quantization, local RAG pipelines, and building agentic workflows using the Model Context Protocol (MCP).

I have built AI-powered CRM systems, deployed custom YOLOv11 food-detection models to TensorFlow Lite, and fine-tuned ResNet-50 models for 3D MRI classification. I enjoy bridging end-to-end AI solutions—from model training and fine-tuning to scalable API deployment—with modern full-stack architectures.

Education

Sikkim Manipal Institute of Technology

B.Tech in Information Technology (IT)

2019 -- 2023

Experience

2023 – Present

Auctus Digitech

Full-Stack & AI Engineer
  • Engineered production-ready CRM systems and AI chatbots leveraging local open-weights LLMs (Gemma/Qwen via Ollama) and agentic workflows for 5 clients, reducing API operational costs by up to 60% while improving lead conversion rates by ~30%.
  • Developed personalized email generation systems integrating LLMs and RAG with internal CRM databases to automate tailored communication workflows, boosting engagement rates by 40%.
  • Designed and integrated secure, scalable REST APIs handling 10K+ daily requests, connecting ML workflows with client-facing web interfaces.
  • Optimized and deployed 4+ production web applications using React, Vite, and TypeScript, slashing page load times by ~40% through code splitting and lazy loading optimizations.
Python LLMs LangChain LangGraph FastAPI React TypeScript PostgreSQL Docker

Projects

Void Drop - Serverless P2P File Sharing Portal
Void Drop – Serverless P2P File Sharing Portal
React Vite WebRTC IPv6 / ICE Service Workers Web Audio API Tailwind CSS
  • Built a serverless, zero-knowledge peer-to-peer file sharing application using raw WebRTC data channels for direct browser-to-browser encrypted transfers.
  • Implemented client-side backpressure window controls and a Service Worker-based ReadableStream interceptor to pipe binary chunks directly to the browser's download manager, avoiding memory overflows.
  • Added an IPv6 direct-connect mode that bypasses Carrier-Grade NAT (common on Indian ISPs) where IPv4 STUN hole-punching fails, by restricting ICE candidates to the chosen address family.
  • Shipped a one-click theme toggle between a clean, plain-language modern dark UI and the original animated Cyberpunk 2077 HUD — live ICE path scanner, telemetry feeds and procedural sound.
Launch Void Drop P2P Portal
PrivaSizer - Privacy-First In-Browser Media Suite
PrivaSizer – Privacy-First In-Browser Media Suite
React TypeScript Vite Web Workers Comlink pdf-lib exifr Tailwind CSS
  • Built a privacy-first, fully in-browser media suite in React & TypeScript where files never leave the device, organised behind a studio picker spanning five distinct tools.
  • Shipped image optimize/resize/adjust/watermark, video trim·convert·GIF, an EXIF/GPS metadata inspector, a redaction tool (blur·pixelate·black-out), and a PDF studio (merge, split, rotate, images→PDF).
  • Ran heavy image work in an OffscreenCanvas Web Worker via Comlink for a responsive UI; video & redaction use MediaRecorder + canvas, PDFs use pdf-lib, and metadata is parsed with exifr.
  • Rebuilt the interface from scratch as a modern, gradient-rich dark UI with an easy-to-spot EZ/PRO toggle that scales the number of controls for casual vs. power users.
Launch PrivaSizer Media Suite
CalCal - AI-Powered Fitness & Nutrition App
CalCal – AI-Powered Fitness & Nutrition App
React Native TensorFlow Lite Python YOLOv11 FastAPI Node.js
  • Fine-tuned a custom PyTorch YOLOv11 segmentation model on 40,000+ food images (~87% mAP), porting it to ONNX and then to TensorFlow Lite for edge deployment with 300ms on-device inference.
  • Integrated a YOLOv8 object detection model to identify known reference objects, enabling real-time scale calibration for accurate food volume and quantity estimation.
  • Architected and published the full-stack mobile application to the Google Play Store, spanning 3 microservices (FastAPI, Node.js, TFLite inference) and integrating the CalorieNinjas API.
Download CalCal AI Fitness App on Google Play Store
Local Repo Sentinel Dashboard
Local Repo Sentinel (WIP)
LangGraph FastMCP Python/TS Qdrant FastAPI React Arize Phoenix
  • Engineering a full-stack, autonomous codebase debugger using a stateful LangGraph orchestration loop with persistent memory checkpoints (SqliteSaver) to self-heal code in response to test traceback errors.
  • Building a custom Model Context Protocol (MCP) server in Python/TypeScript to expose tools for safe filesystem traversal, AST dependency analysis, and local test suite execution.
  • Implementing an advanced retrieval pipeline utilizing semantic query-rewriting, hybrid search (Qdrant + BM25), and local Cross-Encoder reranking (BAAI/bge-reranker-large).
  • Exposing real-time agent execution transitions via FastAPI WebSockets to a React-based control panel displaying a live pulsing LangGraph State Visualizer and OpenTelemetry/Arize Phoenix observability.
Development in Progress
ASD Detection
Deep Learning for ASD Detection
B.Tech Major Project Medical Research Python TensorFlow ResNet-50
  • Developed and optimized deep learning models to detect Autism Spectrum Disorder (ASD) using cortical thickness measurements for computational medical research.
  • Processed 3D MRI scans from the ABIDE dataset, applying data augmentation to improve model generalization.
  • Implemented and fine-tuned a ResNet-50 architecture, achieving a 79% classification accuracy and reducing model loss by 25% through iterative hyperparameter tuning.
Read Deep Learning ASD Detection Case Study
Project Normandy
Project Normandy
Python PyGame
  • Built a retro arcade-style space game with optimized spaceship controls and adaptive enemy AI, improving gameplay fluidity by 40%.
  • Designed an intuitive GUI with animated backgrounds and sprites, enhancing user engagement by 60%.
Explore Project Normandy Source Code on GitHub

Skills

Languages

Python
TypeScript
JavaScript
Kotlin
Java
SQL

AI / ML & Deep Learning

LLMs
NLP
Generative AI
Computer Vision
Model Fine-Tuning
Model Quantization
Object Detection & Segmentation
RAG
Agentic Workflows
LangChain
LangGraph
MCP
Ragas

Frameworks & Libraries

PyTorch
TensorFlow (Lite)
Keras
OpenCV
HuggingFace
NumPy
Pandas
FastAPI
React
React Native
Node.js

LLM Tools & Platforms

Ollama
llama.cpp
LM Studio
Continue
Cline
Arize Phoenix
LangSmith
OpenTelemetry

Databases & Vector Stores

PostgreSQL
MySQL
Qdrant
ChromaDB
DynamoDB

Tools & Infrastructure

Git
Docker
Kubernetes
Linux
AWS (EC2)
Microservices
Vite
ComfyUI
Unity (AR/VR)

Spoken Languages

English (Fluent)
Hindi (Fluent)
Bengali (Native)
Assamese (Native)

Certifications

AWS Cloud Training Certificate

Udemy (Backspace Academy - Paul Coady)

Dec 2021

Comprehensive course covering AWS cloud concepts, core services, architecture patterns, deployment, and security.

View Certificate
AR/VR Training Certification

Internshala Trainings

2021

Completed training program covering Unity engine, C# scripting, and core principles of Augmented and Virtual Reality development.

View Certificate

Contact

Get in Touch

Resume

You can view or download the complete PDF version of my resume reflecting my latest projects and skills.

Download Rajarshi Das AI Engineer Resume PDF