Paul Jablonski

Overview

A real-time granular synthesis VST3 plugin built with the JUCE C++ framework, fully compatible with DAWs like Ableton Live and FL Studio. The plugin deconstructs incoming audio into micro-sound grains and resynthesizes them into new textures by controlling grain size, timing, pitch, density, and randomness — originally scoped for FPGA deployment, then pivoted to software to prioritize flexibility and rapid iteration.

Plugin Interface

How It Works

Audio is continuously fed into a circular buffer, which is sliced into individual grains at configurable sizes and intervals. Each grain is independently pitch-shifted, time-stretched, and shaped with a Hann window envelope to eliminate clicks and discontinuities at grain boundaries.

Grains are dynamically layered according to density and randomization parameters, producing textures ranging from shimmering pads to glitched artifacts. A dry/wet blend control mixes the resynthesized output back against the original signal, and all parameters are fully automatable within the DAW environment.

Overview

A data-conservative Android security system that eliminates storage waste by isolating and storing only human subjects — not entire frames. People are detected via background subtraction against a calibration image, cropped to tight bounding boxes, and assigned unique identity labels through Euclidean color distance comparison, enabling cross-reference across footage with minimal storage overhead.

Detection Examples

DSP Pipeline

Live frames are bitmask-differenced against a calibration image to isolate foreground subjects. Canny edge detection outlines each subject, and a custom flood-fill algorithm groups edges by orientation. Groups are reduced to their average pixel position, then density-based clustering forms tight bounding box proposals around each person.

Each proposal is analyzed with a k-means implementation to extract the six dominant colors from subjects. These color vectors are compared against stored entries via Euclidean distance to either assign an existing label or register a new identity.

Fully manual implementations of flood-fill, DB-scan clustering, and k-means color analysis — no external CV libraries beyond OpenCV for Canny.

Overview

A mouse-shaped robotic cat toy built from scratch with a custom PCB, CAD-designed enclosure, and Arduino firmware. Beyond basic movement, the toy features a distance-measuring laser for collision avoidance, a vibration sensor that detects when the cat catches it, and a microphone-and-speaker system for voice command interaction — allowing the user to locate or activate the toy hands-free. Received an Honorable Mention award at UIUC, placing in the top 7 among nearly 50 groups.

Design & Build

Hardware Design

The toy is driven by two NEMA 17 stepper motors via A4988 drivers, with two SG90 servos handling the animated tail and treat-dispensing latch at the rear. An ATmega328p microcontroller serves as the central state machine, coordinating sensor inputs and motor outputs across four subsystems: sensors, microcontroller, motors and outputs, and power.

Power is supplied by an 11.1V lithium-ion battery regulated down through three AZ1117 voltage regulators, with a CYT1091 button for clean power switching. Audio output is driven by a PAM8302 amplifier paired with a JSM 2.5mm speaker for voice feedback and sound effects.

Functionality Demos

Overview

A GCP-hosted music sharing platform where users register accounts, log in, and create posts rating songs and artists from their listening history. Song, artist, and user data is sourced via the Spotify API and stored in a mySQL database on GCP, with a React and Node.js frontend serving pages for posts, artists, music, and friends.

Stack & Architecture

The backend establishes a Node.js connection to the GCP mySQL database, populating the frontend with live data across all pages. SQL stored procedures handle login and signup validation against stored credentials, while triggers maintain data integrity across post insertions, deletions, and edits.

Initial database population was automated via a Python script using the Spotify API, generating CSV files that seeded the mySQL tables with real song and artist data. The frontend is composed of TSX components styled with CSS, with each page — Posts, Artists, Music, and Friends — driven by parameterized SQL queries.

Pages

Overview

A recreation of the classic flash game Soccer Heads, implemented in SystemVerilog and C on a Microblaze FPGA board. Two players compete in a timed 2D soccer match — one controlled via keyboard, one by an onboard AI — with real-time physics, sprite animation, scorekeeping, and full VGA/HDMI display output.

Gameplay & Demo

System Architecture

A top-level SystemVerilog module ties together the VGA controller, color mapper, physics ball module, player and AI modules, FSM, scoreboard, hex display drivers, and Microblaze block design.

SPI protocols connect peripheral devices via a MAX3421e interface, and ROM-based block memory IPs store sprite and palette data via generated COE files.

Overview

A handwritten character recognition app built with a linear SVM model trained on the EMNIST dataset. Input images are convolved with edge and line detection kernels before classification, achieving over 85% accuracy. A PyQt desktop UI allows users to draw characters directly on a canvas, which are preprocessed, convolved, and classified in real time.

Model & Pipeline

Training data is passed through 2D convolution with identifying kernels that detect edges and lines within 28×28 input images. The resulting feature representations are used to fit a linear SVM, which is then exported and imported into the UI application. User-drawn input is preprocessed and recentered before being fed through the same convolution pipeline and classified by the model.

Interface