INSIDE UPSTART / CHAT

Type a question. Get the answer. Seconds, not days.

In Upstart, Chat turns a plain-English question into an answer, a chart, a table, and a citation of exactly how it was computed — scoped to what you’re allowed to see.

Request a demo → How it works ↓
upstart.education / chat
You · Show me proficiency by grade, this year vs last — for Lincoln HS.
✓ understanding ✓ schema filter ✓ access controls applied ✓ query validated ✓ run ✓ logged
Grade'25'26
954%68%
1061%75%
1149%62%
1257%71%
deeper by school which IEP goals correlate? trend over 3 years
· · ·
You · Which IEP goals correlate with this gap?
✓ analyzing IEP-goal corpus ✓ cross-referencing assessment results ✓ 2.4s

→3 IEP goal categories show statistically significant correlation with the math-proficiency gap:

reading.comprehension.grade-level r = 0.71
math.problem-solving.multi-step r = 0.64
executive-function.task-completion r = 0.52
→47 students in grades 11–12 have at least one of these goals AND scored below proficient on the latest MCAS Math.
show that student list add intervention overlay export to Excel

[ SWAPPABLE PLACEHOLDER — real Chat screenshot drops in here ]

HOW IT WORKS

Question to answer in three steps.

01
Ask in plain English
Type your question, or @-mention a student or school. Nickname-aware — "Mike" finds "Michael."
02
Watch it work, honestly
Live status shows every real step: understanding → looking up people → selecting data → validating → applying access controls → running → logging.
03
Get a real answer
A written explanation, the right chart (auto-selected from 15 types), and a sortable table. Then three follow-up chips to keep exploring.
CAPABILITIES

Built for the way educators actually ask questions.

15 chart types
Bar, line, heatmap, scatter, metric card, and more — auto-selected for the shape of your data.
Multi-turn conversations
Up to 10 turns. Full conversation history. Resume where you left off — on your phone, in your office, the next morning.
Follow-up chips
Click a suggested next question — the next query submits itself.
@-mention people
Mention a student or a school. Names are decrypted in-memory, scoped to what you’re allowed to see.
Cross-highlighting
Hover a bar to highlight its row in the table; click a cell to filter the chart. One-click chart download. Fullscreen mode.
Honest guardrails
Clarification cards when ambiguous; access-denied cards that tell you what you can ask. Groups of fewer than 5 students are automatically suppressed.
THE AI IS UNTRUSTED · BY DESIGN

Every other tool sends a question to a model. We treat the model as a brilliant but unreliable intern.

Its output passes through five independent defense layers before a single row reaches you — and the AI never sees the actual data, only aggregates and your own words.

1
Schema filtering — AI only sees authorized tables & columns.
2
Query validation — no writes, no cross-tenant, no system tables.
3
Access-filter injection — role & scope appended after AI generation.
4
Database enforcement — Postgres Row-Level Security, final fail-safe.
5
Result post-processing — groups < 5 students suppressed.
FROM THE PRODUCT, NOT THE MARKETING

Defensible commitments, from shipped code.

Student data is never used to train AI models. AWS Bedrock discards every request the moment the answer is generated — nothing retained, Anthropic never sees it.
Five layers of defense between the model’s SQL and the database.
The model refuses to guess. If the data isn’t there, it says so.

Try Chat on your district’s questions.

Tell us a little about your district and we’ll walk through it on your data.

Request a demo →