Local vs cloud: PhD-level analytics on federal data - in hours, on your Mac.

Note: This session is a public method demo that compared multiple models. The shipping product is turbolapper - AFM (Apple Intelligence on-device; data stays on your Mac). See requirements and Technology.

Multi-series analytics in hours

Homeownership costs, migration waves, obesity and murder rates by state - in one session on a Mac. Work that used to mean weeks of analyst time (load the series, join the grains, build the charts, check the math, chase the second and third question) can land in hours, with the dataset staying on the device and every extra question free on the turbolapper side of the bill.

That is Local Analytics: turbolapper is the analyst method so serious work can run on your Mac. The shipping product (turbolapper - AFM) uses Apple Intelligence only - this write-up still documents a multi-model method demo.

Summary

• Data: Public housing, migration, health, and crime series - a multi-topic report teams often wait weeks for.
• Speed: Once the series are local, follow-ups cost almost nothing - so you actually ask the second and third question.
• Local vs cloud: Same harness, side by side - Claude Haiku (cloud baseline) vs Qwen 3.6 and other local models, plus the upcoming Apple Foundation Model on-device.
• Detail: Local models generate SQL, charts, and citations - and the harness catches soft math (“nearly 80%” when the answer is 77%).
• Next: Watch the session · Join the macOS waitlist

What Local Analytics means here

Cloud AI made answers feel instant - then billed curiosity by the token, and often asked for your data to leave the building.

With turbolapper:

1. Data stays on your Mac - explore federal or private series; the shipping app keeps data on your Mac.
2. Method demo (this session) compared cloud Haiku vs local models including Apple Foundation Model - research harness, not the App Store config.
3. The harness does the heavy analyst work - SQL, charts, citations, confidence - so on-device analysis can still produce a report you can defend.
4. Questions are free to stack - flat monthly app pricing; the second migration cut, the crime overlay, the “what if rates fall?” chart - none of them open a new meter from us.

What we put on the Mac

Bruce and Tim opened a live dashboard of public topic cards (source, last published, next update) and dug into:

• Housing - Zillow Home Value Index + Freddie Mac 30-year rates
• Migration - Census international and domestic flows
• Health & safety - CDC obesity; state murder rates per 100k

Same motion as the retail prep and NYC taxi sessions: open data anyone can follow, method you can reuse on proprietary tables later.

Local vs cloud: the models in the harness

This session is a local vs cloud bake-off on real charts, not a slide about privacy.

A year ago this class of work defaulted to frontier cloud APIs. In the session, multiple local options clear the harness - which is how private, flat-cost curiosity becomes real.

When “nearly 80%” fails

The on-device Apple model rounded a share to “nearly 80%” while the calculation - and the other models - said 77%. The harness failed the soft answer.

Local Analytics is not about hoping the laptop model “sounds smart.” You test models until the data is right, with citations and exact figures - the same standard you would demand from a human analyst who had weeks.

What the session found

Housing: rates ate the budget

Typical ownership cost in the session cut: about $1,871/month - roughly $784 above late 2021. Of that jump, about $604 was the rate piece of the mortgage; the rest tracked home value / principal. Once prices leveled, further cost pressure was mostly rates.

Because the series are local, the natural next questions - what if rates fall? how does this look by region? - do not need a new budget request.

Migration: a cliff, then the usual roads

International migration estimates in the session fell on the order of ~90% (from multi-million annual levels toward a few hundred thousand early estimates). Domestically, California and New York still lose large absolute numbers of residents; Texas, Florida, and the Carolinas keep showing up as destinations. Small states can look dramatic in percent terms even when big states move more people in absolute counts - the kind of nuance you only catch if follow-ups are cheap.

Health, crime, and why people still move

Colorado looked strong on obesity (~25% in the cut shown). Murder rates per 100k vary hard by state; composition (metro vs rural) matters. Some leave-California destinations look safer on the map; others that gain migrants (Carolinas, Florida, Texas) can show higher murder rates than California or New York. The free follow-up is the point: jobs, taxes, and opportunity often beat a single safety stat - and you can check that in the same session without another cloud invoice.

How to run this yourself

1. Load public series (or your own) onto the Mac.
2. Keep source + publish date on every topic card.
3. Run the harness: SQL → chart → citation → confidence.
4. Bench local vs cloud on the same question - pick the model that earns the number.
5. Stack follow-ups while everything stays on-device.
6. When the data is proprietary, use the same motion - that is the product.

FAQ

What is this session actually demonstrating?

Local vs cloud analytics on real federal and market series - and how fast a full multi-topic report can come together when the analyst method runs on your Mac. Watch: YouTube · write-ups live next to other demos on Community.

Which models did you compare - local vs cloud?

Claude Haiku as the cloud baseline; Qwen 3.6 and other local models for on-device SQL/charts; Apple Foundation Model on-device as the upcoming Apple path. The harness grades exact math and citations, not just prose. More architecture: Technology.

Why “hours instead of weeks”?

Traditional multi-series work means waiting on tickets, warehouse queues, and rationed follow-ups. With data local and questions unmetered by turbolapper, the same team can load, chart, challenge, and re-cut in a single afternoon - the episode is a live example of that pace.

Do extra questions cost more?

Plans are flat monthly by capacity and prep depth, so you can keep asking. With Apple Intelligence on your Mac, analysis runs with the work. Pricing · Economics

How do I run this on my own Mac?

Join the waitlist for the turbolapper app, watch the session, then try the same method on public data first. Scale story: NYC taxi. Trust/prep story: retail prep.

Watch and go further

• Watch: Team Data Crunch · federal data session
• Related: Retail prep · NYC taxi scale · Community
• Product: Waitlist · Technology · Pricing