gradcub logogradcub
Contact sales Book a demo
Documentation/The Brain
CASE STUDY · ENGINEERING

One nervous system for an autonomous internship platform

The Brain is the decision layer of the platform. Every autonomous choice made about a student, a project, a meeting, or an agent passes through it, and every one of those choices is recorded, versioned, and replayable.

Engineering team·August 2026·8 min read

"Why did the system send this student a nudge at 2:17 in the morning last Tuesday?"

For a long stretch, we could not answer that question. This is the story of building a platform that can.

The problem

Autonomous behaviour had grown by accretion. A missed intro meeting was handled in one place, a quiet student in another, a model outage somewhere else again. Each handler decided and acted in the same breath, so there was no shared memory across them, no shared policy, and no trail to follow afterwards.

The cost was not only auditability. The same event was treated differently depending on which module happened to catch it, thresholds could not be adjusted without shipping code, and the system re-learned each student's patterns from scratch every day.

BEFORE
Each automation decided and acted in the same place.
The same event handled differently by different modules.
No trail explaining why an action was taken.
Thresholds could only change with a deploy.
AFTER
Producers report; one layer decides.
One rule per kind of event, applied everywhere.
Every decision recorded with its cause and outcome.
Thresholds adjustable while the cohort is running.

How it works, at a high level

Four stages, in order. Producers no longer decide anything; they report that something happened and return.

01
Report
Anything worth reacting to is reported as an event. No decision, no side effect.
02
Decide
A versioned rule reads the event and what the system already remembers, and proposes an action.
03
Deliver
The action is scheduled rather than fired, and survives restarts and deploys.
04
Remember
The event, the decision, and the outcome are kept together as one record.

Three ideas that mattered

RULES AS CODE
Every decision carries the version of the rule that made it

Rules live as reviewed, tested code rather than settings in a dashboard, and each recorded decision is stamped with the version that was live at the time. Six months later, an audit can reconstruct which policy applied, without guessing. Thresholds inside a rule can still be tuned while a cohort is running.

BELIEVABLE DELIVERY
Actions wait, on purpose

A nudge that arrives in forty milliseconds reads as a bot. One that arrives a few minutes later reads as someone noticing. Each action is assigned a delivery profile that governs how long it holds before it fires, with anything urgent exempt and sent immediately.

instant snappy human deliberative business hours
MEMORY THAT COMPOUNDS
The system stops re-learning the same student every day

A running summary is kept per subject, so decisions take previous weeks into account instead of only the last event. It feeds staff-facing reports and never surfaces to the student, which keeps the programme from feeling like it is grading them in the background.

One rule worth naming

A missed first meeting can be rescheduled without limit. No strike, no penalty, no effect on standing; the student's guide is only told once it has happened enough times to be worth a human conversation. A first-week hurdle should not decide the outcome of an internship, and a policy layer is the right place to say so out loud.

Every rule of this kind is written once, in one place, and applies the same way to every student in every cohort. That is the point of separating the decision from the thing that triggered it.

What it unlocks

Answer "why did the system do that?" for any past action.
Add a new autonomous behaviour in well under an hour.
Adjust how strict the system is mid-term, without a release.
Guarantee no reported event is silently dropped.
Retire or replace an agent that keeps producing rejected work.
Give guides a per-student picture that improves over the month.
ON SCOPE

This note describes the shape of the system, not its internals. Policy contents, thresholds, signal taxonomy, data model, and operational tooling are not published. If you are evaluating gradcub for a cohort and need more detail, we cover it under NDA.

Talk to us about a pilot →