Why a local AI model?
Every AI feature in Sightline runs on your computer. That is not a privacy feature added afterward. It is the decision that lets a school-based practitioner install the thing at all.
Nothing is disclosed, so there is nothing to authorize.
We moved the hard part into the app.
Getting capable drafting to run with no server behind it is a genuinely hard engineering problem. We took it on so that adoption would be easy. The model is tuned for this work, the analysis pipelines structure the record before it writes, and all of it runs offline.
The choice costs us something real: we ship a model with the app, and it has to earn its keep on hardware we don’t control. What it buys you is a tool you can install this afternoon and still justify in a records meeting.
- Observation record
- Structured analysis
- Local model
- Draft for review
←The app, the model, updates
✕Student data: nothing, ever
The cloud shortcut moves the risk to you.
Cloud report tools offer a sign-up-and-paste path that skips the approvals. The requirement doesn’t disappear when a vendor routes around it. It lands on the person who pasted, and it is worth reading a compliance page the way your district would.
- 1A vendor’s compliance page describes the vendor’s controls. It tells you how they store and handle data. It is not a decision anyone made about your students.
- 2Only a district can authorize a vendor to receive student records. That authority sits with the district, not with the practitioner and not with the vendor. Many states also require the agreement to be in writing, with specified terms.
- 3A personal subscription or a free trial carries no such authorization. The account is yours, the upload is yours, and so is the answer when someone asks who approved it.
A local model ends the question instead of answering it. There is no disclosure to approve, on the trial or ever after.
Every claim traces back to the record.
The AI does not reason from a prompt. It writes from a structured analysis of the record you coded, and what it writes stays attached to the evidence behind it.
The practitioner stays responsible. Checking the language against the record, and signing what you send, remain professional judgment.
Diego left his seat six times across the three observations1, most often during independent writing. Episodes averaged 41 seconds2 and clustered early in work periods. In the peer scan, his on-task rate ran 22 points below the class median3.
Evidence behind this draft
- 1Frequency count · 3 sessions · Oct 12 to 26
- 2Duration record · 12 episodes timed
- 3Momentary peer comparison · Oct 19
Summaries checked against the numbers
A summary’s claims are validated against the session’s computed statistics before you read it. A figure that does not match the record does not reach the page.
Function-analysis hypotheses
From a single session or across the whole case, drawn from the coded record rather than from your description of it.
Cross-session narratives
Trends across a case in prose, descriptive or intervention-focused, built from the sessions you actually ran.
Drafting in the write-up workspace
Refine a selection against a before-and-after diff you accept or reject, continue a paragraph, or draw a transition. Every AI edit is logged.
Citations back to the evidence
Generated prose keeps links to the sessions and measures behind it, so checking the language against the record is a click, not a hunt.
On-device transcription
Voice notes from the iPhone companion are transcribed on your machine and land in the session log with their timestamps.
What it deliberately doesn’t do.
A tool that writes about children should be specific about its limits. Some of these are design decisions we would not trade.
It is not a chatbot you brief.
There is no window where you describe a student and ask for a report. The AI reads the structured record you coded. If the observation was not taken, there is nothing to write from.
The evidence analysis is not AI at all.
Contradiction and convergence checks across sessions are deterministic software. They compute the same answer every time, and we would rather say so than sell them as intelligence.
It does not decide anything.
Eligibility, function, and intervention are professional judgments. Sightline drafts and analyzes; the practitioner checks the language against the record and signs what goes out.
It does not transmit your work.
No prompt, no draft, and no student record is sent anywhere for AI. The app ships one model provider and it runs on your computer.
What it takes to run.
No server, no API key, no account. The practical requirements are a computer you already have and one download.
- Where it runs
- On your computer. macOS on Apple Silicon or Intel, and Windows.
- Network needed
- None for AI. The model downloads once during setup, then it works offline.
- Setup
- A two-step wizard: start the local engine, run a connection test. Re-runnable from settings.
- Included in
- The Pro plan. Recording, results, and export need neither AI nor a Pro license.