Projects
Projects where I turn AI ideas into working applications, from face recognition to local LLM apps.
FaceMetrics-AI
Facial Biometrics System
A facial recognition and biometrics platform with 1:1 verification and 1:N identification, dynamic threshold calibration, ISO/IEC-compliant evaluation metrics and an LLM-powered diagnostic assistant.
Key Features
1:1 Verification
Compares a probe face against a claimed enrolled template, with decision thresholds that update dynamically in the interface.
1:N Identification
Searches the whole gallery and returns ranked candidate matches, with 93.5% Rank-1 accuracy on LFW.
Auto-Orientation & Alignment
MediaPipe detection with eye-line affine alignment and rotation recovery, which brought the EER on a 90° rotated test set from 25.4% back to 0.97%.
Biometric Evaluation Engine
Computes ROC, DET and CMC curves with EER and Rank-1/Rank-5 accuracy, following ISO/IEC 19795 testing guidelines.
LLM Diagnostic Assistant
A Gemini-powered assistant that reviews the metrics, explains FAR/FRR trade-offs and helps troubleshoot results.
PDF Report Export
Exports structured evaluation reports from the assistant as print-ready A4 PDFs.
Tech Stack
AI Chatbot & FastAPI Backend
Local LLM chatbot with real-time streaming
A full-stack AI chatbot that runs the Qwen 2.5 0.5B model locally behind a FastAPI backend, with real-time token streaming and a glassmorphism web interface.
Key Features
Local Model Serving
Loads and serves Qwen2.5-0.5B-Instruct with HuggingFace Transformers directly from a FastAPI backend.
Real-Time Token Streaming
Streams each token to the browser as it is generated, using Server-Sent Events.
Session Conversation Memory
Remembers what was discussed earlier in the current session, so replies keep context.
Two Response Modes
A /stream endpoint for token-by-token output and a /generate endpoint for complete responses.
Glassmorphism Interface
A responsive vanilla HTML/CSS/JS interface with a glass-style design and animated background.
Runs on Modest Hardware
Works on a standard CPU, with faster inference on a local GPU or Apple Silicon.
Tech Stack
Healthcare Analytics Dashboard
ECG/EEG Signal & Medical Image Analysis
A full-stack web application for processing and visualizing biomedical time-series data such as ECG and EEG, together with clinical images, with a React frontend and a Python/Flask backend.
Key Features
Real-Time Signal Analysis
Processes and visualizes biomedical time-series such as ECG and EEG signals in the browser.
Spectrum Processing
Spectral analysis of medical time-series data, built on SciPy.
Medical Image Processing
Tools to analyze and process clinical images directly in the app, using OpenCV and Pillow.
Interactive Visualizations
Dynamic charts built with Chart.js and a React interface animated with Framer Motion.
MySQL Data Storage
Stores patient and health data in a MySQL database connected through SQLAlchemy.
Modular REST API
A Flask API organized into separate route modules for analysis, data and image processing.
Tech Stack
Text-to-Image Retrieval API
Semantic Image Search with CLIP & FastAPI
A text-to-image search engine that uses OpenAI's CLIP model and ChromaDB to find images in a local dataset from natural language descriptions, with a FastAPI backend and a glassmorphism web interface.
Key Features
Semantic Search
Finds images from natural-language descriptions using CLIP embeddings, with no tags or captions needed.
Persistent Vector Database
Stores image embeddings in ChromaDB for fast similarity queries that survive restarts.
One-Time Dataset Indexing
A dedicated script encodes the whole image folder into the vector store, so searching stays quick.
Validated REST API
A FastAPI backend with Pydantic request validation and CORS support for easy integration.
Glassmorphism Web UI
A responsive search interface built with vanilla HTML, CSS and JavaScript.
Collection Statistics
A /stats endpoint reports on the indexed collection.