Project Portfolio

ACI Central Intelligence

Keywords: NLP, LLM, Self-host Small Language Models, MLOps

ACI Central Intelligence

ACI Central Intelligence is platform for self-hosted LLMs to be used across the businesses of ACI. It's also used heavily with its developer platforms to enable ACI agentic workflow. The system is enabled for use with low-resource environment. For example, a 32B Qwen model was served using only one NVIDIA RTX 4090 (24GB) by sharding possible model layers to CPU. As of now, different models of 14B (Phi-4), 12B (Gemma-3) are served in the system with quantization to run on a single GPU without any sharding for faster inference. The system also utilizes a novel algorithm for on-demand model loading, ensuring no GPU overhead during no-demand periods.

Key Features of ACI Central Intelligence:

  • Incredibly low-latency by utilizing C++ for model laod and inference (llama-cpp).
  • Custom llama-cpp pipeline for multimodal inference.
  • Automated on-demand model load and unload algorithm.
  • Streamlit-based simple chat frontend.
  • Quantization of models for use with lower GPU overhead.
  • Languages: Python 3.10, C++, Docker

ACI Insight Explorer

Keywords: NLP, LLM, Insight Generator, Business Insights

ACI Insight Explorer is a highly appreciated project for business data analysis at ACI. The system uses AI and LLMs (powered by ACI Central Intelligence) to gather insights, detect anomalies and suggest improvements without looking at actual business data. To keep the privacy of sensitive business data, the system uses novel algorithms to provide mathematical and descriptive summaries of each columns instead of individual data points. The self-hosted LLM then provides code for mathematical operations and visualizations to generate suggestions and anomalies in the business data.

Key Features of ACI Insight Generator:

  • Incredibly low-latency by utilizing ACI Central Intelligence.
  • Data privacy ensuring by not sharing sensitive datapoints and using data summary instead.
  • Languages: Python 3.10, C++, Docker
LLM Generated Visualisation, Anomaly Detection
LLM Generated Insight Over Data
Summarization of Business Data

ACI HR Management Development Planning

Keywords: NLP, LLM, MLOps

Management Development Planning

ACI HR Management Development Planning (MDP) is a part of ACI HR Agentic Hub initiative. This initiative takes effort for automating HR workflow (recruitment, assessment, onboarding, and progress of an employee). The MDP project generates development plans for existing employees using AI, powered by ACI Central Intelligence.

Key Features of ACI Central Intelligence:

  • Incredibly low-latency for usage with self-hosted LLM in ACI Central Intelligence.
  • Custom generation pipeline for ACI employees.
  • React-based interactive frontend.
  • Language: Python 3.10, React, Docker.

ACI SpeechHub

Keywords: NLP, ASR, Diarization, YouTube to Doc

Try SpeechHub

ACI SpeechHub is a combination of several audio based applications developed for meeting automation, call centre automation and several other business usage. The entire ML backend was developed by me integrating models for speech recognition, speaker diarization, dialogue summarization and keyword detection.

Key Features of ACI SpeechHub:

  • Incredibly low-latency (approximately 50s for 1 hour of audio) Audio Transcription in Bengali and English languages.
  • Speaker Diarization and dialogue-style conversation generation from audio data.
  • Summarization of entire conversation.
  • Mentioned keywords detection and frequency count.
  • YouTube to PDF generation.
ACI SpeechHub

Key Technologies used in ACI SpeechHub:

  • Transformer-based (Whisper-Medium) Automatic Speech Recognition (ASR) system.

  • Flash attention 2.0 based Insanely Fast Whisper technology used for incredibly low latency transcription.

  • Integration with PyAnnote based speaker diarization for dialogue style conversation generation.

  • BERT based dialogue summarization technology.

  • Levenstein distance based Keyword detection system.

  • Integration of FastAPI based devOps system and SQL based database system for seamless usage.

  • Backend language: Python 3.9, PyTorch

The code for this project can't be made public for security reasons, but can be provided with proper check. Contact me via Email if you needed.

CLICK HERE TO USE ACI SPEECHHUB

Face Recognition and Live Attendance System

Keywords: Deep Learning, Face Recognition

The face recognition based live attendance system uses IP camera in the office compound to detect faces of employees and generate automated attendance system. Each registered employee's face is detected and then the entry and exit time is written in the database. The database further can calculate total office hour spent by an employee from the recorded system.

Key technology used:

  • FaceNet architecture based face recognition system.
  • Anti spoofing system for detecting false positives.
  • Qdrant based vector database system for storing and later matching facial embeddings of employees.
  • Backend language: Python 3.8
20240606_155226
A fun image while developing face recognition system

Document Keyword Extraction for Supply Chain PO & PI Documents

Keywords: Deep Learning, Document Extraction, Prompt Engineering

Sample Response

Important information extraction from supply chain PO and PI documents using Google's multimodal "Gemini" model with proper prompt engineering. The problem statement was to fill in business forms by extracting several important information such as "Importer/Exporter Name", "Bank Details" etc. from supply chain documents.

Key technology used:

  • Google’s Gemini 1.5 Flash API based service.
  • Prompt Engineering by experimenting with several types of prompts.
  • Backend Language: Python 3.8

ACI Project HitCounter

Keywords: DevOps, Logging System. Gradio

The ACI Project HitCounter is a project developed for documentation of each project usage among the businesses. This project interrupts each API requests in the ACI servers and creates a log file. Later, the frontend reads the logfile for a comprehensive overview of each project usage with count of total project usage as well as count of each endpoint requests.

Key technology used:

  • API middleware to interrupt API requests of each projects and create logs.
  • SSH technology for server-to-server communication.
  • Gradio for frontend.
  • Backend Language: Python 3.9
ACI MIS Project HitCounter

Bangla NLP Toolkit - PyPi Package

Keywords: NLP, Deep Learning, PyPi

Package Link

BanglaNLPToolkit is a package for several classic NLP text preprocessing and augmentations for Bangla NLP tasks.

Key features:

  • Bangla Text Normalization.
    • Bangla text unicode normalization for text preprocessing using bnunicodenormalizer and csebuetnlp/normalizer.
    • Removal of punctuations or replacement of punctuations with desired sign as user desires.
  • Bangla Punctuation
    • Add punctuations to Bangla texts with no punctuations.
    • Uses deep learning based Named Entity Recognition models for accurate punctuation addition.
  • Bangla Text Augmentation
    • Text augmentation techniques for generating similar but different texts for augmenting Bangla dataset.
    • Uses paraphrasing, cross translation and masked word prediction algorithms for augmented text generation.
  • Simple Bangla Tokenizer
    • Robust simple word level tokenizer for Bangla texts.

Click here to learn more about the usage of the package, redirects to GitHub repository and documentation.

Bengali Book Genre Classification

Keywords: NLP, Deep Learning, Classification

Kaggle Link

This was a competition arranged by CUET ETE DAY 2023 on Kaggle. The problem was to classify books on 7 different genres from reviews/summaries collected from the internet. The dataset contained a mixture of Bengali and English reviews of books on all 7 of the genres. My solution to this competition involved heavy data processing and experimenting with different models to achieve the best score. I succesfully achieved the 1st Runner Up position in the competition and 3rd position in the private Leaderboard (the 1st position on the leaderboard was a fake submission from a deleted account).

Resume Classification Using Natural Language Processing

Keywords: NLP, Deep Learning, Classification

GitHub Repository

Building a resume classification system using transformer based NLP systems. The model was trained to classify resumes in several common predefined classes. I fine-tuned the XLNet base model to build the system. This was an independent project of my own in an attempt to satisfy my curiosity of NLP. The details of the project and my training process are available in the GitHub repository mentioned below.

Health Monitoring System with PPG & ECG Signal Using IoT

Keywords: Health Monitor, Early Health Warning, IoT

Project Report

By utilizing PPG and ECG sensor data, real time monitoring health condition of individuals using Arduino. The sensor data was then utilized to the cloud for early detection of diseases and warning the individual. This work is an academic project evaluated for the Biomedical Signal Processing & Instrumentation laboratory course I took during my final year of undergraduate. I got an 'A+' for successful completion of the project. A detailed report can be found in the following link.

Traffic Speed Detection System with Image Processing

Keywords: Traffic Detection, Image Processing

GitHub Repository

Detection of vehicles in traffic footage and count velocity using YOLO V2.0 based image processing system. This was my first ever project in the field of Machine Learning. Me and my group used YOLO V2.0 library to build and run the system in MATLAB environment. The project was an academic requirement for Digital Signal Processing laboratory course during my third year of undergraduate. I got an 'A+' in this course as well.

Morse Code Generator with FPGA

Keywords: Morse Code Generator, FPGA, Verilog HDL

By utilizing FPGA hardware and Verilog HDL, a system was built to convert given input to their respective Morse code. This project was an academic requirement for the Digital Logic Design Laboratory course of the 3rd year of undergraduate. I got an 'A+' in the course and the project received great appreciation from the course teachers.

Solar Powered Environmental Monitoring System with Arduino UNO

Keywords: Solar, Temperature Humidity and Gas Level Monitor, Arduino

Video Demonstration

Arduino based environmental monitoring system. The system uses solar to power itself and run automatically to detect temperature, humidity and air quality of the environment. This project was an academic requirement for the Microprocessor Laboratory course of my final year of undergraduate.

Implementation of an MQTT Based Home Automation System

Keywords: IoT, Smart Home, MQT

Project Report

IoT based system to controll home appliances with smartphone. The project was demonstrated in proteus simulation environment along with a practical implementation in one of our group-mate's own home. This project was an academic requirement.

Analysis of Three Phase Circuit with MATLAB

Keywords: Circuit Analysis, Three Phase Circuit, MATLAB Code

Project Report

Building a system which can solve complex electrical circuits and three phase circuit systems using MATLAB and build a user interface for general use. The complex circuit solver was the real challenge. The system can take input of circuit of any combination or any kind and still solve it using MATLAB codes and generate results. The user interface makes it easier to navigate. This was my first academic project, a requirement for the Numerical Methods Laboratory of my second year of undergraduate degree.