I'm building EvalPi to help AI product managers evaluate and optimize agents with a single instruction.

The planned workflow uses PRDs, evaluation criteria, and connected agent APIs to generate tests, analyze failures, propose improvements, and verify results through regression checks.

Explore EvalPi

I'm building SleepClaw to help people make sense of a night's sleep. Device records can show when you slept, while your own experience adds context about how the night felt and what might have affected it.

The current desktop app supports self-reported sleep and Apple Health XML/ZIP imports. With a connected model, a Pi agent asks focused follow-up questions and queries relevant time windows. Calculated metrics stay separate from AI interpretation; the report brings together the evidence, limitations, and one practical action.

The first milestone is a useful single-night investigation that people can pause and return to. Personal baselines and longer-term action tracking are planned for later stages.

I built grill-powers after using Superpowers on product work and watching product questions mix with technical ones. Without an engineering background, it is hard to tell which answers are product decisions. Each new technical option can reopen scope, so the requirement list grows and rarely shrinks.

The skill joins Grill Me and Superpowers with hard stage gates. Grill Me settles product decisions one at a time and freezes an approved product spec. Superpowers owns technical design, the plan, build, review, and fresh checks. If design or coding finds a real product change, work pauses, returns to Grill Me, and re-walks approve → spec → plan before resume.

I want it to serve product managers and solo builders who need Agents for delivery without giving up product control when scope shifts mid-build.

I built desktop-organizer from my own need to clean a Windows Desktop filled with documents, screenshots, installers, archives, and code projects. Manual sorting required repeated decisions, while direct automation introduced the risks of misclassification, overwrites, and unintended file operations.

The workflow is inspect, plan, approve, execute, and verify. Its default mode only prints a dry-run; moving requires explicit approval. It does not delete or overwrite, conflicts stop execution early, shortcuts and system files stay in place, uncertain items enter a review inbox, and all work remains local.

I want it to serve heavy Windows Desktop users and develop into a personalized classification system where people can keep their own taxonomy, rules, and exception lists.

I built clip-md after repeatedly feeling friction when saving useful AI content. ChatGPT, Claude, and other tools often produce answers worth keeping, and I wanted those answers to become part of my local knowledge base. The old flow required an editor, a new file, a destination, a filename, and a final save. That sequence was long enough to interrupt the thought I was trying to preserve.

I reduced the core interaction to one small action: copy text, use the control that appears near the pointer, and save it directly as local Markdown. Automatic naming is another deliberate choice. The app extracts keywords and uses a fallback chain when extraction is weak, removing one more decision while keeping files readable and searchable.

I want clip-md to serve frequent AI users and knowledge workers. Its direction is deliberately lightweight: improve reliability, platform fit, signed distribution, and the quiet details that let a desktop tool stay useful for a long time.

I built quant-research to turn investment ideas into hypotheses that data can challenge. A strategy can sound reasonable and still rely on one favorable period. I want to see which data it used, how it formed signals, where returns came from, and how much survives when windows, parameters, and models change.

I pay particular attention to overfitting and fragmented research chains. Parameter searches can produce attractive in-sample results, while code, data, and temporary outputs slowly lose their connection. I keep weak factors, failed models, and degraded strategies, and I evaluate costs, attribution, drawdown, and out-of-sample behavior alongside return. A failed experiment still tells me which path lacks durable evidence.

I want quant-research to become a reusable framework for snapshots, hypotheses, signals, backtests, validation, and recorded findings. New markets and models can then enter the same structure, making personal research cumulative, comparable, and reviewable.

I built crypto-trading-research around a question that directly affects my own decisions: after fees, slippage, turnover, and attention are counted, is long-term BTC holding already the lowest-cost and lowest-effort strategy? Active trading and machine learning can look sophisticated, but complexity alone does not prove additional value.

I use long-term holding as the persistent baseline, then test classic indicators, statistical signals, and machine-learning models against it. Costs are explicit because frequent turnover can erase a paper advantage. Windows and fee assumptions remain visible, and model metrics still need to become real strategy returns. MA, RSI, Bollinger Bands, and XGBoost experiments remain in the repository even when their results are weak.

I want the project to become a BTC decision evidence base that keeps testing the same question across new data and market regimes. Every conclusion stays attached to its sample window, cost assumption, and validation method so another learner can review the whole path.

I built ecommerce-user-analysis to understand how transaction records become meaningful customer segments. Revenue, order count, and average order value describe the whole business, but they compress differences between loyal customers, drifting customers, promising newcomers, and people who need a different strategy.

I use RFM to create an interpretable business segmentation, then K-Means to inspect the natural structure in the data. Together they preserve a clear operating language while adding a second analytical view. The project moves from more than one million historical transactions through cleaning, exploration, segmentation, and visualization, then turns the result into retention, reactivation, and conversion hypotheses.

I want it to develop into an e-commerce growth analytics console that tracks how segments change, which metrics matter for each group, and what growth experiments actually change behavior.

I built ai-chat-analytics because likes, dislikes, retries, and follow-ups do not directly tell an AI product team what to fix first, why users are dissatisfied, or whether an improvement changed the experience. Broad labels such as general quality can locate a problem while still leaving the cause unclear.

I split the work into synthetic data generation, behavior analysis, issue attribution, follow-up signals, and insight summarization. A fixed seed makes the public snapshot reproducible without exposing real conversations. Transparent keyword rules establish an explainable baseline and expose their own limit: 71.2% of current bad cases still fall into a catch-all label, and manual validation remains incomplete.

I want the project to serve AI product leaders and grow into an AI experience monitoring console, with continuous metrics, attribution quality, alerts, and experiment results connected in one place.