I hand-build a private AI assistant on a Raspberry Pi 5. This month it learned to tutor math and plan projects.

Hi, I'm Jacob. I build PHNTM One by hand, one unit at a time, from my desk in the United States.

I wanted an assistant that was actually mine: on my desk, remembering what I told it last week, without sending a word of it to someone else's computer. So I built one, and then I kept going. This post is what it is today, what it costs, and where it is slow, with the numbers.

PHNTM One is a private AI assistant that lives on your desk. It is a Raspberry Pi 5 (8 GB) inside a 10.1-inch 1920×1200 touchscreen case, assembled by hand, with a speaker and a tap-to-talk USB mic. In Private mode the language model, memory search, speech recognition and speech output all run on the device. Unplug your router and it keeps working.

It costs $799 once. There is no PHNTM cloud account and no required subscription for local use.

PHNTM Tutor teaches a learner how to find the answer instead of handing it over. It has 52 linked skills across five subjects, including 38 mathematics skills with 76 worked examples, from arithmetic, fractions and percentages to supported linear equations, coordinate graphs and selected geometry. Lessons run in English, Italian and Spanish. Grading is done by exact arithmetic and supported symbolic checks; the AI model never grades, and there is no cloud fallback. A learner tries a problem, gets guided questions, explains why a method works, then proves it alone on a different problem. Type a problem, write it by hand on the screen, or photograph a worksheet with your own USB webcam (not included); you confirm the recognised text before the lesson starts.

Parent view is for the adult beside the learner: every learner at a glance, progress, due reviews, suggested support and a printable weekly report. It shows what was solved without help in Tutor and what needed help. Practice plans of up to 12 lessons, eight learning paths across five subjects, practice assignments, and separate student and teacher sign-ins.

PHNTM Missions turns "I want to build a birdhouse" into a plan of milestones and ordered steps with the materials and tools each step needs and what counts as done. In Private mode the plan is written on the device by the local model and checked by fixed rules before you accept it; you review and edit it first. Then it keeps an honest record: a step ticked on your word is "reported", not "verified"; a photo you confirm raises the step's grade; and hazard steps (mains electricity, power tools, heights) cannot be marked done without evidence.

Raspberry Pi 5, 8 GB RAM · 10.1" 1920×1200 IPS, 5-point touch · 256 GB microSD · built-in speaker · USB tap-to-talk microphone · mini wireless keyboard with touchpad · 27 W USB-C power supply (100–240 V, US plug; adapter included on most international orders) · Gigabit Ethernet and Wi-Fi 2.4/5 GHz · Debian 13.

A 4-billion-parameter model is slower and less capable than frontier cloud models. On the review unit on October 30, 2026, warm answers had medians of 19 s to first words and 28 s to the complete answer; the first uncached question of a conversation took 87 s and 94 s. Longer answers or a model reload can take over a minute. The model can be confidently wrong and OCR can miss text, so check cited sources. There is no clock battery: after an offline power loss, check the time before relying on reminders.

$799 once, plus applicable tax. Hand-built to order and burned in for 72 hours before it ships. Ships within 7–10 business days after payment. Free US shipping; international orders pay a flat international shipping charge shown before payment (currently $99), with duties and import taxes paid on delivery. 30-day returns, 1-year limited warranty.

It is for people who want the finished, private thing on a dedicated box: notebook people, project people, homeschool families and anyone who would rather keep their train of thought in the room. If you need instant answers or frontier-level reasoning, a cloud model is a better fit. If you have a good GPU and like tinkering, run Ollama yourself.

Website, parts list, transcripts and manual: https://www.phntmcore.com Press kit: https://www.phntmcore.com/press