Some links may earn Mr1Tech a commission at no extra cost to you. We won’t invent hardware benchmarks or “hours of battery” claims — test your own machine before a deadline.
Cloud AI is useless when the power cuts, fibre dies, or data runs out mid-deadline. This is a resilience playbook: what’s actually doable offline, where Ollama fits, and how to talk to clients without promising magic. Joburg outages are the stress test — the same habits help anywhere the network flakes. For SA schedules, use the official Eskom load shedding portal, not a WhatsApp rumour.
Failure modes (name them before they hit)
| Failure | What breaks | What still works if you prepared |
|---|---|---|
| Power cut | Desktop/Wi‑Fi router | Laptop on battery; phone hotspot later |
| Fibre / ISP down | Cloud ChatGPT/Claude/Jasper | Local models; queued drafts on disk |
| Mobile data capped | Heavy cloud uploads | Pre-downloaded models; offline notes |
| Platform outage | Vendor API | Local rewrite + email the client a delay note |
If your whole business is “open ChatGPT in the browser,” one outage means you deliver nothing. Fix the workflow, not just the subscription.
What’s Realistic When You Run AI Offline
Doable offline (with prep):
- Rewriting proposals, emails, and blog drafts you already have on disk
- Summarising files you saved locally
- Coding assistance against a local repo (if your machine can run the model)
- Continuing Notion/docs you synced earlier (live sync needs network)
Not realistic offline:
- Fresh web research with citations
- Image gens that only exist in the cloud
- “I’ll just download a 70B model” on a weak laptop mid-outage
- Guaranteeing cloud-quality every time on a small local model
Tell clients the truth: offline mode is continuity, not a full cloud clone.
Ollama: Offline ChatGPT Alternative (Local LLM)
Ollama is a common way to run open models locally (and related workflows). Local model use is typically free of cloud token fees; your costs are your hardware, electricity, and time.
Practical rules:
- Install and pull a small model on a good-power day — not during Stage whatever.
- Confirm it runs a sample proposal rewrite before you sell “outage-proof” delivery.
- Keep prompts and client briefs in plain files (
proposals/,briefs/) so you’re not trapped in a tab. - Treat local output like cloud output: human edit before send.
We won’t list fake VRAM requirements — check Ollama’s model cards / your machine docs.
Hardware honesty
- Light laptop: may only handle small models; fine for short rewrites, slow or impossible for big contexts.
- Stronger machine / more RAM: more headroom — still test, don’t assume.
- Desktop only: pair with a UPS if outages are common, or accept that desktop work stops when power stops.
- Don’t buy hardware on hype for one client — upgrade when outages repeatedly cost you money.
No affiliate push for a specific laptop here; if Ads adds hardware later, keep it careful and disclosed.
Outage Playbook: Before, During, After
Before (prep)
- Eskom / municipal schedule bookmarked (loadshedding.eskom.co.za for Eskom-linked schedules)
- Ollama (or chosen local stack) installed; one small model pulled
- Client briefs + drafts saved to disk (and optional offline-capable notes app)
- Cloud work queued: anything that must be cloud done in the next open window
- Client delay template saved (SMS/WhatsApp/email)
During (power or network down)
- Switch to local rewrite / offline files only
- No promises that need live web search
- If battery is low: ship a partial + honest ETA, don’t burn the last 5% on tinkering
After (back online)
- Sync Notion/Drive if you use them (Notion pricing if you’re choosing a plan — free tier may be enough for solo notes)
- Run cloud-only steps (images, deep research, final polish)
- Send the finished piece + short note if anything slipped
Privacy bonus (side benefit, not the main sell)
Local models keep prompts on your machine unless you upload them. Useful for sensitive client wording — still follow your contract and platform rules. Privacy is a bonus of the outage setup, not a reason to skip NDAs or common sense.
When local isn’t enough
Escalate to cloud (when you’re back online) if you need:
- Accurate live facts / prices / news
- Heavier reasoning than your local model can do
- Image or multimodal work you only have in cloud tools
- Client-mandated stack (e.g. they require ChatGPT Enterprise output)
Your job is delivery, not ideology.
Soft next step
Want the outage checklist on one page you can print? That’s the sheet (~$2–3 / R30–50). The article above is still the full free playbook.
One Response