All projects
A distributed real-time data processing system that ingests taxi location
streams, processes events with low latency, and pushes live results to a dashboard.
- Focus
- Real-time ingestion and aggregation of high-volume location data with low latency, designed
for monitoring moving systems.
- Architecture
- Events are ingested via Kafka, processed in an Apache Storm topology, and key state is
stored in Redis for fast access.
- Delivery
- Aggregated results are broadcast to a live dashboard using WebSockets for immediate visual
feedback.
- Tech
- Apache Kafka, Apache Storm, Redis, Docker, WebSockets.
A small x86_64 operating system written in Rust that boots under QEMU and
implements core OS concepts such as paging, interrupts, a RAM-based filesystem, and a basic
interactive shell.
- Focus
- Build fundamental OS building blocks from scratch: boot setup, memory management, interrupt
handling, and command execution.
- Architecture
- Boots inside QEMU, initializes paging and interrupt tables, uses a TAR-backed RAM disk at
root, and provides a simple interactive shell with tab completion.
- Runtime
- Experimental support for executing WASM applications that render to the framebuffer and
return control back to the OS.
- Tech
- Rust, x86_64, QEMU, paging, interrupts, WASM, framebuffer.
A small-scale legal AI assistant: upload contracts, extract structured data, run playbook reviews, and ask
questions with exact source citations.
- Focus
- Grounded answers for legal documents, where every extracted value and answer is backed by a quote,
page number, and location hint.
- Pipeline
- PDF/DOCX parsing, overlapping chunking with page tracking, embeddings stored in ChromaDB, and
document metadata in SQLite.
- Features
- Field extraction (document × field table), playbook review (rule → ok / deviation / missing),
and multi-document Q&A with citations.
- Tech
- Python, FastAPI, ChromaDB, SQLite, PyMuPDF, provider-agnostic LLM adapter (OpenAI, Ollama, Gemini).
Fault-tolerant distributed message broker with dynamic cluster coordination
and leader election.
- Architecture
- Implemented exchange and queue abstractions for reliable inter-broker communication.
- Discovery
- Dynamic broker registration via custom DNS-based service discovery.
- Resilience
- Leader election implemented using Raft, Ring, and Bully algorithms.
- Tech
- Java, distributed coordination algorithms, custom networking layer.
Longitudinal analysis tooling for dependency update behavior across Java,
Python, and JavaScript ecosystems.
- Focus
- Measured propagation speed and adoption patterns of dependency updates over time.
- Parsing
- Processed Maven, Gradle, Python, and JavaScript manifests across full commit histories.
- Output
- Generated structured reports highlighting ecosystem-level update dynamics.
- Tech
- Python, GitPython, manifest parsing, data analysis pipelines.
A Nextcloud app that renders a folder's README directly at the top of the Files view, similar to GitLab.
- Focus
- Lightweight, native-feeling README rendering inside Nextcloud Files.
- Features
- Case-insensitive README detection, read-only task lists, syntax highlighting, and auto-refresh
on save, rename, or delete via the Nextcloud event bus.
- Compatibility
- Supports Nextcloud 26–34 using the Files header API with a legacy registration fallback.
- Tech
- JavaScript, PHP, webpack, highlight.js, Nextcloud app framework.
A native GTK4/libadwaita app for AirDrop between a Linux laptop and nearby Apple devices on Fedora GNOME.
- Focus
- Bring the Omdrop AirDrop panel from Omarchy to GNOME as a clean, stateless front end, without
touching the underlying protocol.
- Architecture
- The window drives the omdrop command and reads its JSON output; privileged radio
operations run only through pkexec.
- Packaging
- Install script builds pinned RPMs from source and installs them with dnf, with dry-run modes
and a careful, reversible uninstall.
- Tech
- Python, GTK4, libadwaita, RPM, Shell.
Interactive GPU-based 3D volume rendering for MRI and CT datasets.
- Rendering
- Implemented direct volume rendering for real-time exploration of volumetric data.
- Pipeline
- Three.js and GLSL shaders for GPU rendering, D3.js for analytical support views.
- Integration
- Combined visualization and analytics workflows for structured medical data analysis.
- Tech
- Three.js, GLSL, D3.js, WebGL.
An audio fingerprinting and retrieval system inspired by the classic Shazam
approach, designed to match short or noisy audio clips to a reference database.
- Focus
- Fast, robust retrieval using spectral fingerprint extraction and hash-based matching.
- Approach
- Converts audio to time-frequency features that are resilient to distortion or noise.
An advanced route planner built as a team project, with a cross-platform app for web, iOS, and Android.
- Focus
- Software architecture, collaboration, and implementation quality in a larger monorepo with
separate backend and frontend.
- Architecture
- Java backend with Python helpers for activity data (Strava, Garmin Connect), an Ionic/Angular
frontend packaged natively via Capacitor.
- Delivery
- Push notifications for PWAs (Web Push / VAPID) and native Android (Firebase Cloud Messaging),
GitLab CI pipeline, and Kubernetes deployment.
- Tech
- Java, Maven, Angular, Ionic, Capacitor, TypeScript, Firebase, Kubernetes, GitLab CI.
A desktop GUI for the Codex CLI that combines AI-assisted coding with an embedded terminal and a built-in
versioning workflow.
- Focus
- Make AI-driven code changes visible and reversible: see what changed, review it, and roll back in
one click.
- Features
- File explorer with changed-file highlighting, near real-time diff viewer, integrated Git
snapshots, history, and revert.
- Architecture
- Electron main process handles IPC, Git, and the PTY; a React UI talks to it through a safe
preload bridge.
- Tech
- Electron, React, Vite, xterm.js, node-pty, Git.
A small onion routing network where requests travel through multiple relay nodes, each peeling off one layer
of encryption.
- Focus
- Anonymous communication: no single relay knows both the origin and the destination of a request.
- Architecture
- Directory service for node discovery, relay nodes, an originator with CLI and web interface,
and a sample destination service.
- Crypto
- Hybrid layered encryption with RSA-OAEP for key exchange and AES-GCM for payloads.
- Tech
- Python, cryptography, Docker Compose, JavaScript.
A full node for the Kerma blockchain protocol that communicates with peers, validates objects, and maintains
chain state.
- Focus
- Implement the core of a cryptocurrency node: peer-to-peer messaging, object exchange, and
validation.
- Architecture
- Asynchronous networking with asyncio, peer management, object storage, and a mempool for
pending transactions.
- Validation
- Blocks are validated even when referenced transactions are still missing: unknown objects are
fetched from peers with timeouts.
- Tech
- Python, asyncio, Docker, GitHub Actions.
A knowledge graph of hockey players and careers, enriched with Wikidata and used for career continuation
prediction.
- Focus
- Integrate tabular player statistics into a semantic graph and use it for analytics and
prediction.
- Pipeline
- EliteProspects CSV data converted to RDF/Turtle, enriched with Wikidata/Wikipedia links, photos,
and team history.
- Analysis
- SPARQL exploration in GraphDB and an ML workflow predicting whether a player's career
continues.
- Tech
- Python, RDF/Turtle, SPARQL, GraphDB, Wikidata, machine learning.
Time-series classification of Rocket League trick shots using in-game
telemetry and control signals.
- Dataset
- 297 labeled trick shot sequences used for training and evaluation.
- Method
- Extracted temporal features and statistical aggregates for maneuver classification.
- Approach
- Optimized for practical model performance under constrained data conditions.
- Tech
- Python, scikit-learn, time-series feature engineering.
Web platform for structured peer-to-peer study matching and session
coordination.
- Focus
- Designed matching workflows connecting students by subject, availability, and learning
goals.
- Context
- Full-stack web application with structured user profiles, session scheduling, and matching
logic.
- Result
- Reduced coordination overhead and improved discoverability of study partners.