Building in Public
Every win, every lesson, every experiment documented. The process is the product.
Expanded Bathala beyond project management into a full life dashboard. Added expense tracking, task management, notes, and settings views. Built car project views with EUR/PHP currency toggle. Fixed 12 PTY/WebSocket bugs, hardened data isolation for limited users, added tooltips across all views. Currency preference now lives in localStorage via user profiles.
Complete dashboard redesign — glass UI with multi-theme support, collapsible sidebar, command palette. Added M8 execution oversight with agent dashboard tab for monitoring autonomous Claude workflows. The trading desk now has a proper control surface.
Wired autonomous Claude workflows into the trading loop. The bot now runs agent-driven decision cycles — market analysis, position management, and trade execution as coordinated LLM tasks. Stabilised the full CI suite: eliminated flaky async tests, fixed cooldown defaults, made surveillance and gateway tests deterministic.
Launched Bathala — a self-hosted life dashboard with liquid glass UI. Multi-user system from day one: Mito's project command centre alongside Iony's business dashboard with order management. Built-in terminal via WebSocket, pipeline views, and modular architecture. Split the monolith into clean modules within 24 hours of launch. FastAPI backend, Alpine.js frontend, zero external dependencies.
Six LLM features wired into the trading loop: gate (entry filter), regime classifier, exit adviser, trade journal, strategy selector, and parameter tuner. Added news sentiment analysis scoring headlines as entry signals. Strategy research mode analyses trade history for pattern insights. 12 real (non-mock) LLM integration tests. Addressed 4 CRITICALs and 5 HIGHs from code review.
Added Kraken Futures support with REST-only execution at 5x lower fees. Risk sophistication: Half-Kelly sizing, volatility-adjusted positions, correlation-aware portfolio heat. Multi-strategy framework running mean reversion alongside volume confirmation. Multi-timeframe analysis with 4H primary candles and daily trend filter. Trailing stop re-submission and limit order support.
Built a trading dashboard (FastAPI + React + TradingView charts) deployed at trade.genkaw.com. Added backtesting engine with parameter sweep, walk-forward validation, multi-symbol support, and CSV export. Discord notifications for decisions and trades. Corrected Sortino ratio, profit factor, and equity curve calculations after review.
Replaced SMA/EMA crossover with a volatility breakout strategy. Restart-safe trailing stops with persisted highest_price. Architectural hardening of the strategy branch. Health monitor tests and pre-commit test enforcement hook. The strategy is simpler, more robust, and better suited to crypto's regime-driven volatility.
Complete ground-up build of the Kraken trading system. Five-layer architecture: Exchange, State, Decision, Execution, Control — each with full test suites. Package restructured from flat src/ to proper kraken_llm layout. Ruff linter, GitHub Actions CI, Docker deployment, systemd service configs. From empty repo to production-ready in one sprint.
Tore down and rebuilt the entire Bitvavo trading bot. Migrated from timer-based polling to event-driven WebSocket architecture with SOL as primary asset. Five-layer redesign: exchange, state, decision, execution, control — each audited and tested independently. 150/150 test battery across 50 rounds and 3 suites. Tiered LLM sessions, async Discord, health endpoint. Multiple audit passes fixing 18+ bugs across all layers.
Converted the bot from swing trading to active day-trading. 15-minute sessions, sentinel monitoring, exchange stops. Rules grew from initial set to 41 — covering anti-complacency, missed setup tracking, buy-hold benchmarking, deployment deadline expiry. Added lesson promotion system where insights survive but one-off observations expire. Separate Discord channels for quick alerts vs full analysis.
First build of the Bitvavo trading bot. Shell scripts prep market data, LLM reads files and decides. Support/resistance entries, position limits, trap awareness. Seven core trading principles to prevent beginner mistakes. Systemd timer deployment. Started with Sonnet, switched to Opus within two days for better reasoning on complex market states.
Full repository audit. Fixed 5 correctness bugs, added ARIA tabs pattern to filter components, arrow-key navigation on dropdowns, proper focus management on mobile nav. Restructured content architecture and extracted inline prop types. Added GitHub Actions CI pipeline for PR validation. Standardised colour scales and eliminated DRY violations.
Comprehensive server hardening across every layer. Nginx: brotli + gzip pre-compression, file caching, epoll workers, WordPress proxy cache (HIT latency dropped from 1.2s to 0.3ms). OS: sysctl network hardening, zram swap (2 GB compressed, lz4), TCP Fast Open, kernel BPF hardening. Security: Fail2ban with four jails, Cloudflare Access on Bathala and Filebrowser, UFW with zero inbound rules. Docker: pinned image tags, healthchecks, resource limits, log rotation. Removed TeamViewer (450 MB), cups-browsed (CVE-2024-47176), avahi-daemon.
Moved all five archived experiments to a dedicated /archive page. Each project gets proper real estate—full description, prominent lessons learned, and links to case studies or GitHub repos. The Lab section is now clean: active projects only, with a link to the archive.
Made the site match reality. Eight experiments archived with honest lessons learned. Case study pages preserved with archive banners. Navigation trimmed to essentials. The narrative arc: explore widely, learn deeply, focus ruthlessly.
Final maintenance pass before archiving. Fixed all 236 test failures and 160 errors—green suite. Cleaned dead code, stale docs, dropped unused provider dropdown. Added multi-stage Dockerfile targets. Stax ends in a healthy final state: production-ready code that nobody will run.
Built an RL agent for Command & Conquer Generals. Explored advanced architectures: transformers, mixture of experts, attention mechanisms, RND curiosity for exploration. Fixed 36 bugs across combat, save/load, and particle systems. The architecture exploration was the point—production viability was never the goal.
Built an AI-powered invoice scanning tool with LLM extraction. Auth hardening with middleware, session warnings, upload flow with data integrity. Dashboard restructure with analytics. Reached MVP but shelved—LLM accuracy for structured invoice data isn't reliable enough for solo maintenance.
Completed Phase 3 of the CLI RPG. Chapter 1 branching narrative with player-driven path selection, multi-chapter story progression, complete training system with challenge and sparring feedback. v0.3.8 with Characters section and updated roadmap. Phase 4 planned but project archived—shipping deadline discipline won over scope ambition.
20 of 97