What the Owner Is Working On

This page had not been updated since March 2025. At the time, I was building a browser translation extension, polishing website interactions, and considering a crawler-driven technical weekly. That record is still true, but it no longer describes the work.

The current direction is more coherent: I am connecting products, AI infrastructure, engineering practice, and writing into a system that one person can operate over time.

August 2026 update

StratThread: turning a quant prototype into a product

Most of the work is going into StratThread, known internally as qrp-platform. It is no longer a small tool for viewing prices or running one backtest. The product is being shaped around a complete research loop: market discovery, immutable strategy versions, historical validation, parameter experiments, paper trading, live execution, and notifications.

Recent work has concentrated on a few boundaries:

• Market data now spans OKX crypto, TDX A-shares, and Alpaca US equities. The less visible work is just as important: catalogue ownership, historical backfill, confirmed candles, and recovery when a source fails.
• The global Page Agent can use the current page as context, discover available actions, and receive Chinese or English prompts and tool descriptions from the request language. Its model traffic now routes through llm.geekcode.vip.
• Normal, Pro, and ProMax plans, a seven-day usage window, physical Worker-attempt accounting, and operator controls have reached implementation and regression coverage. The database migration and production release gates are still recorded separately; implemented does not yet mean deployed.
• The market workspace is gaining complete drawing interactions—selection, dragging, colour, width, and deletion—while the Dock, indicators, keyboard paths, and narrow-screen behavior continue to be tightened.

The engineering work behind those features now follows specification-first admission, explicit domain boundaries, recoverable asynchronous jobs, and layered regression. The latest recorded complete Web run contained 152 suites and 981 tests. The immediate unfinished task is to undo an over-broad global visual restyle while keeping the first-use empty-state illustrations that passed review.

Infrastructure: making a one-person operation observable

PostgreSQL, Redis, and ClickHouse now live in a shared Kubernetes baseline. Each application receives its own database, account, and namespace instead of starting another local data stack. New object storage is moving to Cloudflare R2 rather than recreating the retired MinIO service.

OpenLIT now collects LLM traces, token usage, cost, and latency. OTLP ingestion remains private to explicitly authorized cluster workloads, and trace data has a seven-day ClickHouse retention policy. The point is not to accumulate infrastructure. It is to make cost, failure, and recovery boundaries visible enough for one person to operate them.

The same baseline is starting to absorb older products. fortune.geekcode.vip has moved into the unified workspace, restored its PostgreSQL data, and passed production acceptance with two replicas against the shared Redis, LLM gateway, and R2 boundaries. One historical avatar URL remains unresolved because the original image bytes could not be recovered; all new writes use R2. TBTI has also moved into this workspace, while automated release registration is intentionally waiting for an approved repository remote.

GeekCode: turning shipped work into reusable material

The GeekCode website has a clearer job now. It is not a generic developer community or a portfolio. It documents how an AI-native one-person company handles products, engineering, content, and operations in practice.

I rebuilt the brand and documentation experience, with Chinese and English treated as first-class content rather than one language being an afterthought. The latest articles come directly from project work: four guardrails for reliable AI features; a specification, architecture, and changelog loop for stable Vibe Coding; and a practical argument about tests, context cost, regression speed, and iteration. The delivery method behind those articles is also available as the downloadable stable-vibe-coding Skill.

The old crawler-driven technical weekly is still not complete. For now, I would rather publish slower field notes with evidence and limitations than produce summaries merely to maintain a posting frequency.

What comes next

The near-term work is concrete:

1. Finish StratThread’s visual rollback, Agent-plan migration, and production acceptance while continuing to distinguish “implemented” from “deployed.”
2. Keep moving older applications onto explicit recovery, credential, storage, and release boundaries without recreating retired services for convenience.
3. Turn recurring engineering problems into Chinese and English articles and reusable Skills, without inventing a story merely to have something to publish.

Archive: March 2025, week 2

Simple Translate extension

• feat: Finish the professional-enhanced translation capability, including model fine-tuning and corpus preparation.
• feat: Improve the interaction flow with SSE streaming so long translations do not time out.
• pref: Improve the i18n configuration for error messages.
• fix: Refine when the translation button appears, so it is not triggered unexpectedly in certain situations.
• fix: Add a few small guardrails and constraints.
• pref: Improve the Redis data model and add expiration control for stored records.

GeekCode operations

• feat: Add a technical-weekly feature to the website to enrich content and improve SEO (crawler task, API, and frontend).
• feat: Add emoji support to the public docs. 😈
• docs: Expand the independent-development articles. The first two chapters are complete.
• perf: Polish document-anchor and section-highlighting details on the website.
• docs: Add this “what I’m working on” section to build in public and keep a regular record of goals and progress.