Today Eat: a private meal-decision tool
Designed a local-first product path that helps one person decide what to eat without turning into a restaurant finder or delivery platform.
- Scoped the product around one decision and excluded maps, merchants, transactions, and social features.
- Built the core around 96 local candidates, hard constraints, deterministic ranking, and explicit feedback.
- Worked with Agents on product specs, implementation, automated checks, and visual QA; the planned 10-person, 7-day validation is not complete.
- Extended the same product into a verified 25.024-second Remotion launch video workflow.
Jarvis: personal AI knowledge and Agent collaboration system
Connected raw inputs, verification, durable knowledge, publishing, and feedback in a traceable workflow.
- Separated Raw, knowledge, and public output so machine summaries do not silently become facts.
- Operationalized a processing ledger, daily and weekly reviews, publication queues, and human approval gates.
- Completed a real WeChat and Xiaohongshu publishing loop and brought qualitative feedback into the next iteration.
Personal distillation: cross-Agent evidence protocol
Explored how multiple Agents can capture observable behavior without turning isolated conversations into identity claims.
- Separated observable facts, stated reasons, outcomes, candidate inferences, counterevidence, and confidence.
- Verified a structured handoff for Agents without workspace access and preserved source-time conflicts instead of normalizing them.
- Kept the work explicitly experimental because long-term evidence is not available yet.