Ty Dale/Arizona/private AI and creator products

DalesAI, Private AI Infrastructure.

I'm Ty Dale. I build AI systems that do real work: private AI for businesses that want to own their stack, and finished digital products for creators, built from what their audience keeps asking for.

107 GB modelrunning on one Mac
48 of 48agent tasks passed
Private by designlocal models, local data
50 / 50revenue split for creators

01 / For Creators

Your audience keeps asking. I build the answer.

I built an AI system that reads what your audience keeps asking for, checks that people already pay for that kind of product, then builds the product, the store listing, and the revenue split. You only have to share a link.

New in November 2026First product finished50 / 50 split

An inside page of the sample planner: a one page weekly tracker with tables for school days and progress marks
Cover of the sample planner, The 20-Minute Homeschool Day, a printable prep and track planner
fig. 01 · the first finished product: a printable homeschool day planner · real pages, shown as a sample

How it works

  1. It listens

    The system reads the public comments on your videos and counts the questions that keep coming back.

  2. It checks

    It looks for proof that people already pay for that kind of product before anything gets built.

  3. It builds

    It writes and lays out the product and drafts the store page. AI tools do the writing, and a person checks every page.

  4. You share

    I list it on my Whop store with your half already set. You share one link with your audience.

The deal, in plain words

  • I build it and host it. The product lives on my Whop store. It costs you nothing.
  • You get 50% of every sale, after Whop's fees. Whop pays your half into your own Whop account, which is free to open.
  • Your name is used only with your OK. Until you say yes, it stays off the product.
  • You can ask for changes before anything goes live.
  • AI tools do the writing. A person checks every page before you see it.

Where this standsThis is new. The system is built, and the first finished product is the planner shown above.

What it could build for you

Examples only

These are examples of what the system can build. They are not past work. The planner above is the one finished product so far.

  • HomeschoolingA printable day planner. Built: it is the sample above.
  • CookingA four week dinner plan with a shopping list for each week.
  • FitnessA 30 day beginner plan with a daily check off sheet.
  • CraftsA printable project planner with a supply list and cost tracker.
  • Personal financeA first budget workbook: monthly budget, bill calendar, payoff tracker.
  • GardeningA first vegetable garden planner with a planting log.

Got an email from me? Just reply to it.

No email yet? Send me a link to your channel and I'll tell you honestly whether I see a product in it.

02 / Why Private

This is what private AI actually looks like.

Not a thin wrapper around someone else's API. Open models, a model switcher, memory limits, and a test harness, all running on a Mac you can physically point at.

Cloud calls made by my local model server: 0

03 / The Platform

What actually runs here today.

Most "AI consultants" resell someone else's API. I run open models on my own hardware, so when I build for you, I know every layer because I've operated every layer.

A request comes inJARVIS · NIGHT JOBS · CREATOR SYSTEM
OpenAI style APILOCALHOST · PORT 8001
Model switcherLOADS ON DEMAND · IDLE UNLOAD
Qwen 3.8 27BTENSORFOLD
Qwen 3.8 Flash NextMTPLX · 107 GB
Qwen 3.5 9BMTPLX · LIGHT
Apple M5 Max · 128 GB unified memoryONE BIG MODEL IN MEMORY AT A TIME

fig. 02 · how a request reaches a model on my Mac · animation

Models

Three open models and an embedder

Qwen 3.8 Flash Next, 107 GB on disk, is the main brain. A Qwen 3.8 27B runs every nightly job, a small Qwen 3.5 9B takes quick questions, and an embedding model powers search.

Switching

One big model in memory at a time

My own switcher loads a model when a request arrives and unloads it after an hour of quiet, so a 107 GB model and everyday work can share one Mac.

Engines

Chosen by measurement

MTPLX runs Flash Next and TensorFold runs the 27B. Each one won a timed test against other engines before it got the job.

Testing

A 48 task eval harness

Real agent work: edit files, run commands, read code, browse. Graded by checking what the agent actually did, not by asking a model how it went.

Public API

Built in 2026, parked now

I also built an OpenAI compatible gateway with login, usage dashboards, and Stripe billing. Its health check still answers, and this page reads it live, but the models behind it were retired with my old engine, so it is not taking requests.

04 / The Lab

Two months of testing, written down.

I don't pick models by hype. Every model and engine on this Mac had to win a measured test first. Here is what I ran from mid September to mid November 2026, what I decided, and why.

Running today

Checked Nov 16, 2026
  • Qwen 3.8 Flash NextMain brain for my agent. About 177 billion parameters, with about 6 billion used per word.
    MTPLX · 107 GB
  • Qwen 3.8 27B, 4 bitEvery nightly job: study, review, voice notes, the weekly test run.
    TensorFold · 16 GB
  • Qwen 3.5 9BQuick, light questions.
    MTPLX · 8 GB
  • ModernBERT embedSearch by meaning across my agent's skills.
    oMLX · 288 MB

Writing that customers read goes to a Claude model in the cloud. In a blind test on Nov 14, Claude Fable 5.2 beat my local model.

Same model, different engine

Tokens per second, code

Qwen 3.8 Flash Next · Sep 18 to 19

Out of the box13
oMLX, tuned48
MTPLX98

Qwen 3.8 27B · Oct 1 · four of nine setups

Ollama25.2
oMLX33.4
MTPLX56.2
TensorFold69.7

fig. 03 · measured on this Mac · a token is about three quarters of a word

Test log · Sep 15 to Nov 15, 2026

  1. Sep 18

    Tuning a 107 GB model

    Out of the box, Flash Next wrote about 13 tokens per second. Turning on multi token prediction and finding the right draft depth took it to 31 on code. Setting the fans to max, with nothing else changed, took it to 40.

    DecisionKeep the fans up. Heat was costing about 30%.

  2. Sep 19

    Engine showdown

    Same model, two engines. oMLX wrote code at 48 tokens per second, MTPLX at 98. On my 48 task test the scores tied at 47 and 46, but the median task took 15 seconds instead of 42.

    DecisionMoved the main brain to MTPLX: 2.5 times faster on real work.

  3. Sep 19 to 23

    Fitting a huge model on one Mac

    A 107 GB model fills most of a 128 GB machine, so I built a switcher that loads models on demand. Then the lighter 27B crept up to 97 GB in long chats, because its engine sized itself to the whole Mac. With a memory budget, a replay of a real 68,763 token chat peaked at 49 GB.

    DecisionCapped the 27B. First perfect 48 of 48 the same day.

  4. Oct 1

    Nine setups, one winner

    I timed the 27B on nine engine setups. TensorFold with a draft model wrote code at 69.7 tokens per second, against 25.2 on Ollama. A long chat test then caught it growing from 33 to 69 GB, which the short benchmark had hidden. Two cache caps held it flat with no speed loss.

    DecisionNightly jobs moved to TensorFold: 16 minutes for the 48 tasks instead of 38. The old engine stays as a one click backup.

  5. Oct 2

    A guard for the agent

    My agent broke its own model switcher twice by writing settings that do not exist. A guard now checks every edit against the engine's real options and blocks the bad ones.

    DecisionThe agent can change its setup, but only through a check.

  6. Oct 31

    An update, tested before trusted

    Before keeping a new MTPLX version I reran everything: 48 of 48 in 18.5 minutes, and a chat grown to 86,000 tokens with memory flat.

    DecisionKept the update, with the old version saved as a one step rollback.

  7. Nov 5 to 8

    A win that was not real

    The same model on a different engine finished my 48 tasks minutes faster. Before switching I looked for the reason: my test setup was breaking the current engine's cache, so it reread its instructions 35 to 41 times a round instead of 3. I fixed the test and ran both again. They came out level.

    DecisionKept the current engine and deleted the 98.7 GB download. Test runs can no longer change my agent's real memory or skills.

  8. Nov 14

    A blind writing test

    Three writers wrote the same product from the same research: my local model, Claude Opus 5.5, and Claude Fable 5.2. Each draft was printed in the same design and labeled A, B, or C at random. I picked one without knowing who wrote it.

    DecisionFable 5.2 won and wrote the first product. The local model builds and checks. Writing that customers read goes to a Claude model.

What the testing taught me

01

The engine matters more than size or bits

Engine swaps bought 2 to 2.5 times the speed. A 4 bit copy of the 27B matched the 8 bit one on the test at a third of the memory.

02

One or two points is noise

Two identical runs in November finished 6.4 minutes apart. I only trust big gaps and repeated runs.

03

Test long chats, not short benchmarks

Memory problems only show up as a chat grows, so every new engine has to pass a growing chat test.

04

Local models make things up

So I check files and command output, never a model's own summary. It is also why a person checks every page of a product.

05 / What I Build

AI systems for businesses that do real work.

From the first missed call to a full product: this is what I can build and run for your business.

AI receptionists and lead follow up

Missed call text back, appointment booking, and instant lead response for service businesses, so inquiries at 9 PM become customers instead of voicemails.

  • SMS
  • Booking
  • After hours

Private AI assistants

Internal chat and document assistants running on dedicated hardware, for businesses that want AI without sending customer data to a third party cloud.

  • Your documents
  • On premises
  • No cloud

Automation and integration

Connecting AI to the tools you already use: CRMs, calendars, payment systems, email. Webhooks and APIs that just work.

  • CRM
  • Calendar
  • Payments
  • Email

Full product builds

Auth, billing, dashboards, and deployment: the same kind of stack I built for my own API, built for your idea.

  • Auth
  • Billing
  • Dashboards
  • Deploys

06 / About the Builder

Ty Dale, self taught builder.

I learn by building, and I run what I build: a local model server, a home media server, and a personal AI agent that studies every night. I've shipped an app to the App Store, I'm finishing a picture book app for kids, and I design and build websites by hand, like this one. My newest build turns what a creator's audience asks for into a finished product. I'm not a classically trained engineer, and that's the point: I build with modern AI tools the way your business will actually use them.

01

Local AI Systems

Open weight language models served privately on Apple Silicon, with no cloud bill.

  • MLX
  • MTPLX
  • TensorFold
  • Open weight LLMs
02

Home Infrastructure

A self hosted media and photo server, a whole home ad blocker, and secure remote access, all keeping itself running.

  • NAS
  • Jellyfin
  • Immich
  • AdGuard
  • Tailscale
03

iOS Development

Native apps with in app purchases, taken all the way through App Store review.

  • SwiftUI
  • StoreKit
  • Shipped apps
04

Website Development

Custom websites designed, built, and deployed by hand, like this one.

  • Custom design
  • Scroll animation
  • Responsive
  • Cloudflare

07 / Selected Work

Things I've built for myself first.

Keep scrolling

Perika app screen showing how an HRT change affected symptoms
Live on the App Store
01

Perika

A perimenopause and HRT tracker that shows whether a treatment change is actually working.

  • React Native
  • Expo
  • RevenueCat
Jarvis online

Good morning. Night study finished: skills sharpened, notes filed in the vault.

What's on my calendar today?

02

Jarvis

My personal AI agent on Telegram. It runs on my own Mac, studies every night, and gets its skills sharpened by a nightly review.

  • Hermes
  • Local LLMs
  • Obsidian
  • Python
M5 Max128 GB unified memory
Qwen 3.8 Flash Next107 GB · Jarvis brain
Qwen 3.8 27Bnight jobs
ModernBERT embedsearch
03

Local Model Server

Large open models served from one Mac: a 107 GB model with multi token prediction, a lighter 27B for night jobs, and embeddings for search. It powers Jarvis and my creator product system.

  • MTPLX
  • TensorFold
  • oMLX
  • Apple Silicon
Kids book app bookshelf screen
Kids book app reading a page aloud
04

Kids Book App

Native iOS picture books for ages 2 to 4, read aloud by a puppy narrator.

  • SwiftUI
  • StoreKit 2
  • AI narration
Storage7 TB
  • RAID 5 healthy
  • Jellyfin
  • Immich photos
  • Home ad blocker
  • Remote access
  • Nightly backups
  • Battery backup
05

Home Media Server

A 7 TB NAS that runs my home: self hosted Jellyfin and photo backup, a whole home ad blocker, secure remote access, and a watchdog that keeps it all running.

  • Docker
  • Jellyfin
  • Immich
  • AdGuard
  • Tailscale

08 / Demo

Watch a missed call become a booked appointment.

A demo for a business like a day spa: the AI receptionist answers a missed call by text and books the visit. Then the same phone runs a private assistant answering from local files. No human touches either one.

Demo conversation · business and names are illustrative

fig. 04 · missed call text back, as the customer sees it

09 / Work With Me

Have a project? Let's scope it.

If you want slideware, hire a consultancy. If you want a working system from someone who runs his own, tell me what you're trying to automate or build. I'll reply with an honest read on whether it's worth doing, and a fixed price if it is. Creators: the same address works for you.

Replies within one business day