Projects

A selection of projects across embedded systems, data engineering, and applied AI.

Face recognition pipeline using embeddings and MediaPipe face mesh

Face Recognition Pipeline (Dataset Preparation, Image & Video Processing)

Built an end-to-end face recognition system combining embedding-based and landmark/mesh-based approaches. I created a custom dataset by extracting image frames from the NVIDIA “Avatar Fingerprinting” video dataset and organizing them per identity. Using OpenCV, I automated frame extraction and preprocessing, generated per-frame feature embeddings, and averaged them to create a stable identity embedding per person. Recognition was then performed by matching query faces (from images or video streams) against the stored identity embeddings. In parallel, I implemented an alternative recognition method using MediaPipe Face Mesh to evaluate mesh-based features under pose and illumination variations.

  • Dataset: video-to-frame extraction & dataset curation (OpenCV)
  • Embeddings: imgbeddings, per-frame embeddings → averaged identity vector
  • Mesh-based: MediaPipe Face Mesh for landmark-based recognition
  • Processing: image + video pipelines in Python
Ultrasonic rail inspection system

Railway Track Fault Detection (Ultrasonic Signals → CNN Classification)

Worked on an end-to-end workflow for railway fault detection using ultrasonic inspection data. The project covered data acquisition, preprocessing and visualization to validate signal quality, conversion of processed measurements into image representations, and training of CNN-based classifiers for automated defect/state classification.

  • Data: ultrasonic acquisition, preprocessing, visualization
  • Modeling: MATLAB CNN classifiers (AlexNet, ResNet)
  • Tooling: Deep Learning Toolbox, Image Processing Toolbox, Signal Processing Toolbox
Desktop application project

Desktop Image Editor & Bluetooth Transfer to ESP32-S3

Built a Python desktop application that loads and edits images and transmits them wirelessly over Bluetooth to an ESP32-S3 device. The tool supports core image operations such as cropping, resizing, RGB conversion, zooming, combining images, and overlay composition—designed for a smooth operator workflow and reliable embedded-side transfer.

  • UI: PyQt5, Qt Designer
  • Image Processing: OpenCV, Pillow
  • Communication: Bluetooth (SimplePyBLE)
Temperature control project

HOQA Temperature Control System (Raspberry Pi + MAX31865)

Implemented a closed-loop temperature control system using a MAX31865 RTD sensor with a Raspberry Pi. The controller monitors temperature continuously and automatically toggles a heater using configurable thresholds (hysteresis) to maintain a stable operating range—improving safety and preventing frequent switching.

  • Platform: Raspberry Pi
  • Sensor: MAX31865 temperature sensor
  • Control: threshold-based heater control with hysteresis
Amazon product data extraction using Selenium, Pandas, and SQL

Amazon Product Data Extraction (Selenium + Pandas + SQL)

Developed a scalable web data extraction pipeline to collect structured product information from Amazon, including product name, price, product image URL, description, availability status, and sold/stock indicators. The workflow automates navigation and page interactions using Selenium, transforms raw HTML into clean tabular datasets with Pandas, and stores results in a SQL database for querying and analytics.

  • Extraction: Selenium automation (search, pagination, product pages)
  • Data Fields: name, price, image, description, availability, sold/stock info
  • Processing: Pandas cleaning, normalization, CSV export
  • Storage: SQL database schema + insert/update pipelines
UK doctors contact data extraction and enrichment pipeline

UK Doctors Data Collection & Enrichment Pipeline (Web Scraping + Google Maps API)

Built an automated data collection pipeline to aggregate UK doctor and clinic contact details from multiple client-provided web sources. The system extracts structured fields such as email, phone number, address, and practice information using a combination of Selenium for dynamic pages and Requests for fast HTML retrieval. To improve data quality and standardization, the workflow enriches and validates addresses using the Google Maps API, then exports clean datasets to CSV and stores records in a SQL database for reliable querying and updates.

  • Scraping: Python, Selenium, Requests (multi-site extraction)
  • Fields: email, phone, address, practice/clinic details
  • Enrichment: Google Maps API for address validation/standardization
  • Data Handling: Pandas, CSV exports, SQL storage
Germany gas stations data extraction and enrichment pipeline

Germany Gas Stations Data Collection & Location Intelligence

Developed an automated data acquisition pipeline to collect and standardize information on gas stations within selected cities in Germany. The system aggregates data from multiple public and commercial web sources, extracting key attributes such as station name, address, contact details, and operational status. Location data is enriched and validated using the Google Maps API to ensure geographic accuracy and consistency. The processed dataset is cleaned, deduplicated, and stored in structured formats for downstream analytics and reporting.

  • Scraping: Python, Selenium, Requests
  • Data Fields: station name, address, phone, city, availability
  • Geospatial: Google Maps API (geocoding & validation)
  • Data Handling: Pandas, CSV export, SQL database
Automated Weather Data Pipeline using OpenWeatherMap, Selenium, and Docker

Automated Weather Data Pipeline (OpenWeatherMap + Selenium + Docker)

Developed an automated data pipeline that retrieves weather metrics (e.g., min/max temperature, humidity, cloud coverage, wind speed) using the OpenWeatherMap API. The pipeline takes latitude and longitude as input, resolves location context, stores results into structured CSV files, and runs on a scheduled cycle every 60 minutes. The full workflow is containerized for portable, reproducible deployment.

  • Data Collection: Requests, Selenium, OpenWeatherMap API
  • Data Handling: Pandas, CSV
  • Deployment: Docker, scheduled execution
Literature survey automation project

LLM-Assisted Literature Survey Automation (Face De-identification)

Designed and automated a literature survey pipeline focused on face de-identification research. The workflow aggregates results from IEEE, ACM, and ScienceDirect, normalizes records into a unified dataset, removes duplicates, and applies an AI-assisted filtering pipeline to rank and extract the most relevant work. The process reduced ~8,000 papers to a shortlist of 24 highly relevant studies.

  • Data Engineering: collection, preprocessing, deduplication, single clean CSV
  • Automation: Academic Cloud API integration
  • AI: OpenAI API for relevance filtering and summarization