Every repo I've shipped.

A persistent AI built on top of OpenClaw. Lives on the Pi, runs around the clock, picks up threads I have forgotten, drafts replies before I open the laptop. Telegram in and out. A paid model when it is awake on a real task; local Qwen on Ollama for the heartbeat. The Thalamus plugin (next entry) is the routing layer that keeps the context budget under control.

A community plugin for OpenClaw. Replaces "paste the whole transcript when you spawn a sub-agent" with a 3-field reference: packet_id, resolver_key, inline_vector. The receiver resolves only the atoms it needs from a local vector store. Qwen3-Embedding-0.6B runs on Pi 5 CPU via llama.cpp. A FAISS RaBitQ codebook compresses 1024-dim vectors to 128 bytes. MIT licensed. Distributed on ClawHub and npm.

A 9-layer signal engine: trend, momentum, volatility, volume, sentiment, an ML ensemble, Ichimoku, candlestick patterns, and statistical models. Forty-six features pass through Boruta selection and CPCV validation before the stacked ensemble (LightGBM, XGBoost, RF, ExtraTrees with a Ridge meta) makes a call. $100 simulated balance on Binance Testnet. Never connected to real money.

Production-ready Selenium and Gauge specs generated from a small locator JSON. Snapshot-based discovery, locator auto-extraction, a step catalog DSL any QA engineer can extend. Tested against a real consumer-facing booking flow; the happy path runs end to end in under two minutes, with full step snapshots as an audit trail.

After AI-assisted prototypes proved unplayable, this is the deliberate version. Top-down perspective, deep RPG simulation, written from scratch in Unity. A long-game labour of love.

A whimsical TypeScript prototype: agents as pixel sprites moving around a virtual office, each role embodied. Started as a visualisation experiment for multi-agent dialog systems before the Ember pivot.

Java daemon on the Pi exposes GPIO, serial, and sensor data over a local socket. The tablet becomes the display + control surface. Used during the Alcyone build for live monitoring without a monitor.

Telegram bot interface, mixed local model (Ollama qwen3) plus cloud LLM orchestration. Designed as a research-grade testbed for multi-agent dialog systems before Ember.

Gauge-based test automation for an "Example Bank" web app, with the same AI Snapshot + Locator/Test Autogen flow that Pegasus now uses. README contains a full Step Catalog auto-generated from @Step annotations, and "golden rules" for the autogen pipeline (locator JSONs, double-quoted parameters, no improvised steps). Last week's private repo — the bridge between masterAutomation and Pegasus.

Single Java project, three runtime profiles. Locator JSON + step catalog pattern that became the blueprint for Pegasus a year later.

A C# / Python hybrid: ASP.NET API on top of an AutoGen-based Python service running seven agents — Coder, Data_analyst, Engineer, Scientist, Planner, Critic, Executor — coordinated by a ModifiedUserProxyAgent. Built and posted publicly in January 2024, when the word "agentic" was still mostly papers.

A WPF/.NET trading simulator. Backtested simple strategies on historical data, plotted PnL curves. The first time I sat with the trading problem long enough to understand why most retail bots lose money.

Flask front-end over a TA-Lib indicator engine. A workbench for testing technical indicators side-by-side. The codebase later seeded the Alcyone signal layer architecture.

Small C# utility — drop in a Swagger JSON, get a runnable cURL command per endpoint. Useful for QA test bootstrapping and API smoke tests. Built into HealthMonitor.

.NET 8 backend, web dashboard. Upload OpenAPI documents, generate cURL commands, run automated health checks, custom assertions, notifications. Self-hosted alternative to several SaaS offerings. The "$5/month app you don't need to pay for" thesis applied to API monitoring.

A C# / .NET 8 backend for an AI-driven tabletop simulator that respected the Open Game License framework. Three-tier architecture (Data / Services / API), JWT auth, AI-generated storylines via HttpClient. The first serious attempt at the RPG idea Ember now carries.

A more disciplined take on the discovery / locator / test pattern, using a major Turkish e-commerce site as a real-world target. Surfaced the locator-JSON approach that became standard in everything I built after.

A Java app that parses code (classes, methods, fields, annotations) and generates human-readable descriptions using OpenAI GPT-3 and HuggingFace CodeBERT. Customizable description templates, integrates with Java codebases. The first repo where I integrated CodeBERT seriously — and the project that produced the famous "I'm so sorry! I'm so sorry!" infinite-loop screenshot a friend turned into an Instagram story. (A GPT2-style chatbot fed through CodeBERT's tokenizer without fine-tuning will, in fact, just say "I'm so sorry" forever.)

My first real internal project at a previous employer — a UI-driven log extraction tool. Engineers needed to pull specific fields from messy log streams, but most QA folks will not write regex from scratch. parserTest exposes a UI for building extraction rules: you point at the parts you want, name them, choose the format, and the tool generates the regex behind the scenes. The output is a clean, queryable structured log. Years before "log analysis SaaS" had a name, this was already running for the team.

A mobile UI test automation framework I built years before Appium was the standard answer to anything. UI element discovery, gesture replay, assertion DSL — all custom. I tidied the project up and pushed it to GitHub years later as an artefact, but the code itself was running in production tests well before mobile QA tooling consolidated around the current stack.

"Detection of Defects Occurred in Assembly Line by Image Processing" — a parallel master's project alongside the predictive-maintenance thesis. Trained on a balanced dataset of 40,000 cement-surface images, 80/20 train/test split, transfer-learning from open-source VGG16, sweeps over learning rate / node count / dataset size. Final test accuracy 98.3%. The framing: prognostics for cement-based household goods — catch surface cracks before they become field failures. The intellectual bridge between the mechanical engineer I trained as and the QA engineer I became.

"Mahzenler ve Evrenler" is the most literal Turkish translation of "Dungeons & Dragons" you can write. A multi-project C# solution (BLL, DAL, UI, UnitTests, Tester). The architecture is naive in retrospect, but the ambition is the same one Ember now carries: a real RPG, faithful to the source, built in Turkish first. Every RPG I have tried to build since has been an answer to a question this codebase asked. Six years later they made the movie. I had not figured out how to ship things widely yet.

Hosted at koexamapp.msbel.com (msbel.com was already live in 2018, this used a subdomain). Teachers create quizzes from articles auto-fetched via Wired.com RSS, or write their own. Students take exams on a single page; jQuery dynamically updates the page after submission to show correct/incorrect answers in colour — no full reload, an SPA five years before Astro existed. Built as a job interview project. BLL/DAL/UI architecture, rich text editor, full auth.

A single HTML page that serves multiple "pages" via JavaScript and AJAX — no framework, no build step. README from the time: "expandable so data can be fetched with ajax methods. Simple solution for serverless static pages." Nine years later, this site you're reading is the same idea, just with more discipline.

A blog engine written while learning MVC. Crude, earnest, working. Followed by a burst — nine repos in a single day on October 27, 2017, including TruvaTour, TourAutomation, EntityFrameworkExampleApp, UtilityBillsApplication, CSharpDesignPatternExample. The shape of someone teaching himself to ship.