Projects / Case Index

Projects

Concrete problems in human-AI collaboration

01

Agent handoff protocol

SACP

When an agent claims completion, how can it leave a handoff another operator can inspect, judge, and continue?

Ongoing research
receiptevidencevalidatorhandoff
Built
A protocol connecting task receipts, evidence, responsibility boundaries, and the next owner.
Still resolving
Hard model constraints, independent acceptance criteria, and multi-agent coordination remain unresolved.
02

Inspectable execution record

Agent Flight Recorder

Can claims, evidence, gaps, and repairs from an agent run become an externally inspectable record?

Prototype validation
skillauditreceipt hashreflection loop
Built
Integrated SACP-style receipts into a skill and experimented with public receipt-hash anchoring.
Still resolving
The early skill carries too many responsibilities and needs smaller recording, validation, and repair interfaces.
03

Task contract compiler

token-prompt-compiler

How can an ambiguous request become a task contract with explicit goals, boundaries, validation, repair, and stopping rules?

Public repository
objectiveboundariesvalidatorrepairstop rule
Built
Combines prompt study, reflection, and SACP into a reusable compilation and review flow.
Still resolving
Cross-model format differences still need benchmarks and a tighter outcome loop.
04

Markdown-first AI workflow

Solo AI Company OS

How can reusable AI roles, task templates, records, and context stop collaboration from restarting at zero?

Exploring
rolescontextdashboardmemoryhandoff
Built
Organizes AI roles, prompts, context, and handoffs into a Markdown working system.
Still resolving
The architecture still needs long-running evidence for state, storage, and parallel agent collaboration.
05

Web evidence agent

ShadowBuyer

Can inspectable evidence help hotels understand regional differences in prices, taxes, and terms?

Hackathon prototype
multi-agentregional pricingHTMLscreenshotsaudit
Built
Moved from Amazon price differences to Booking hotel data, collection, cleaning, conversion, and evidence packaging.
Still resolving
Collection stability, anti-bot systems, compliance, and pricing definitions still need validation.

Products and short-cycle delivery

06

Language learning assistant

IELTS Assistant

How can materials, writing feedback, and review become a sustainable language-learning assistant?

Live product
OCRwriting coachlearning loop
Built
Turned personal learning friction into an online assistant connecting materials, feedback, and review.
Still resolving
This is a personal learning-system experiment, not a mature education product.
07

36-hour hackathon delivery

Journease

How can an ambiguous idea become a working product under severe time pressure?

Delivery review
problem framingdeliveryreview
Built
Completed framing, collaboration, implementation, and presentation in 36 hours, then reviewed the engineering gaps.
Still resolving
Short-cycle delivery does not prove long-term maintenance or engineering quality.