When life falls apart, information is rarely enough. Stepkin explores a 72-day, AI-assisted pathway that turns advice into appointments kept, paperwork completed, jobs tried and support actually received.
步伴 · 一个公开的 72 天社会复归机制提案。把「我知道该做什么」变成「我今天真正完成了什么」，同时保留真人判断与支持。
A research and design proposal, not a functioning shelter, AI service, or eligibility system.
Today, let's make progress on just one thing.
Speak. Plan. Take the next step.
Government data reveals both movement through shelters and a major blind spot: what happens after someone leaves.
of tracked shelter users exited and were not recorded returning in the following year (2023–24 cohort).
This is not a verified rehousing rate.of those who exited were recorded as entering housing.
54.6% had an unknown destination.of formerly sheltered or unsheltered respondents who regained housing cited new or improved employment income as a contributing factor.
Among people who had regained housing; self-reported and multi-select.Sources: Federal shelter outflow analysis (2025) · Statistics Canada, Exiting homelessness (2025). Different populations and denominators; these percentages cannot be combined into one program success rate.
Start by understanding a person's immediate constraints. Let AI coordinate suitable administrative and daily tasks, and escalate whenever the next step needs expertise, safety assessment, or human discretion.
Employment or income interruption, missed paperwork, daily disorganization, transport and appointment follow-up. The assistant breaks work into small verified actions.
Repeated barriers, literacy needs, disability accommodation, unreliable communication or complex agency decisions. Human support is built into the plan.
Urgent safety concerns, serious medical or mental-health needs, addiction treatment and safeguarding. AI may assist with coordination, never diagnose or decide eligibility.
The immediate problem is income and rent, not a lack of ambition. Progress should begin with essential coverage and housing protection.
It follows up on applications and appointments, prepares the next task and helps the person try again after a setback. An actual professional reviews eligibility and exceptions.
Illustrative pathways; not a medical, legal, housing or benefits decision.
Limit the high-touch coaching window so resources can be measured and reassigned. But never mistake a missed milestone for proof that someone no longer needs food, safe shelter or clinical care.
Safe place, contact device, food, urgent risk check, consent and identity needs.
ID, benefits, appointments, communication, mobility and practical routines.
Discover strengths, paid job trials, training and everyday money skills.
Work progression, housing search and repeated support where tasks fail.
Verified progress; step down, extend or transition to specialized supports.
Start with actual local vacancies, then assess access, suitability, wages and potential. A failed trial should produce information about the next match—not a permanent label of failure.
Example: truck driving may be a meaningful long-term path, but Ontario Class A licensing ordinarily requires at least 103.5 hours of mandatory entry-level training. Windsor–Sarnia truck driving currently has a moderate, not severe-shortage, outlook for 2025–27. Ontario licensing · Job Bank labour outlook
The open design challenge is how to allocate scarce intensive support responsibly without turning an app into a compulsory surveillance gatekeeper.
Payment and evaluation should reflect suitable verified outcomes, adjusted for case complexity. Otherwise, an operator is rewarded for high occupancy or for selecting only the easiest cases.
72 days is an experiment in structured support—not a scientific deadline, not an eviction timer, and not a universal recipe for every person.
Begin with a small voluntary research pilot alongside existing human services. Measure the extra value of AI follow-through relative to usual services while ensuring no participant loses access to standard support.
proposed participants in an initial feasibility pilot.
Design target, not a completed project.days of structured intensive assistance, with follow-up beyond the program.
Not an automatic end to basic services.day outcomes to understand whether changes last.
Housing stability and income where appropriate.Can proactive AI reduce administrative friction? Which work trials are truly accessible? How should a voluntary program respond when someone disengages? Can public-space conditions improve without simply shifting harm elsewhere? These remain hypotheses, not settled results.
Independent original concept, first publicly shared October 2026. Free to discuss and reference with attribution to this project; the creator's personal details are intentionally not displayed. AI-assisted research and drafting. Not affiliated with Cloudflare, municipalities, governments or social-service providers. No sign-ups, case files, recordings or location data are collected by this concept site.