Personal AI assistant · Stella module
Iris is a dedicated AI personal assistant for every manager in your school. She reads your email, drafts replies in your voice, and sends nothing without your approval — while learning your style, your relationships, and your patterns every single day.
Why Iris? In Greek mythology, Iris was the goddess of the rainbow and divine messenger — carrying communications between gods and humans with precision and grace. Clean, international, and easy to pronounce in Italian, English, Irish, and every language across our markets.
#iris-rosanna · personal approval channel
Every manager who uses Iris gets their own dedicated PA instance — not a shared service, not a chatbot anyone can access. Stella spins up a private "job shadow" for each person: an agent that knows their role, their contacts, their writing style, and their recurring patterns.
Iris operates with delegated authority. She can send email on your behalf — but only after you tap ✅. That is a meaningfully higher trust level than any other Stella module, and it is why Iris is designed the way she is: every action visible, every draft reviewable, every send logged.
The more Iris works with someone, the better she gets. Every approved draft she sends without edit reinforces her model of that person. Every edit teaches her something new. Over time, the gap between what she drafts and what you would have written closes — until the approval tap becomes genuinely effortless.
Either a dedicated PA address (e.g. rosanna.pa@school.ie) or monitoring the manager's own inbox via IMAP alongside them — transparent, auditable, always visible.
Role, contacts, writing style, recurring topics, relationship map — built from the sent folder and continuously updated from live inbox and Stella context.
One Zulip thread (or Telegram chat) exclusively for that manager. Every draft, every notification, every briefing arrives there — and only there.
Iris knows your school because she inherits Stella's knowledge layer — policies, timetables, group records, room availability, pricing, staff contacts.
Iris monitors the inbox continuously and acts on what she finds — drafting, classifying, and briefing — all before the manager opens their laptop in the morning.
Action needed · FYI only · Ignore safely. Each classification is logged. The manager sees a triage, not a flood. Iris learns which calls she gets right and adjusts.
Using your sent folder history, your relationship with this contact, and Stella's organisational context, Iris drafts a reply that sounds like you wrote it — because in every meaningful sense, you taught her how.
Draft + original email + research summary arrive in your private Zulip thread. One tap to send. One tap to edit. Nothing goes without your sign-off.
"You have 7 unread. 2 need action: Lorenzo Bianchi re July group and Ana re room 3 change. 3 are FYI. 2 can safely wait." Five bullets. Every morning. Ready before you arrive.
When an email arrives from a contact who has a group arriving next Monday, Iris already knows. Her draft reflects that context — before you have to explain it to her.
Few-shot style: every edit you make to a draft is stored in Iris's per-user memory. Over weeks, her drafts require less editing. Over months, the approval tap becomes a formality.
Beyond autonomous monitoring, Iris responds to direct requests through your approval channel — pulling from Stella, CRM, and calendar to act immediately.
Iris pulls the group record from Stella CRM, identifies who needs to be notified, drafts the coordination emails, and sends them all to your approval channel — one tap per email.
Iris looks up Ana's email, drafts a short message in your voice, sends it to approval. Tap ✅. Done. Total time: 4 seconds.
A structured five-bullet briefing arrives within seconds — sender, topic, urgency, and whether a draft is already waiting.
Iris queries the Stella CRM, your email history, and any outstanding context — returning a structured summary of the relationship before you reply.
Iris reviews outstanding threads, identifies non-responses, and queues follow-up drafts in your approval channel — batched, reviewable, ready to send.
Delivered before 8am: unread count, action items, FYI items, and anything Iris flagged as time-sensitive overnight.
Within minutes of an email arriving, Iris has read it, classified it, cross-referenced it with Stella, drafted a reply, and sent it to your approval channel — waiting for you.
Anything that looks urgent — a complaint, a visa question, a group change with less than 48 hours notice — is flagged immediately, not batched into the morning briefing.
Every 20 minutes, Iris runs a quiet background update — compressing new inbox and calendar signals into her per-user knowledge base, so her context never goes stale.
Every Iris instance begins with a one-time ingestion of the manager's sent folder — the purest available signal of how that person communicates. From there, learning never stops.
On setup, Iris connects to the manager's email via IMAP and pulls every message ever sent. Signatures, quoted text, and forwarded content are stripped away. What remains is pure voice — every word that person chose to write, to every person they chose to write it to, in every context they have faced over years of work.
This corpus is embedded into a private Qdrant namespace under iris:{user_id}:sent — chunked, searchable, owned exclusively by that Iris instance.
From this corpus Iris learns: writing style (formal or casual, sentence length, sign-off patterns, emoji use or deliberate absence) · relationship map (who they email most, how they address different people — "Hi James" vs "Dear Prof. Murphy") · decision patterns (what they approve, what they delegate, what they escalate) · recurring topics and how they are typically handled.
Alongside the personal knowledge base, every Iris instance pulls from the shared Stella namespace — the organisational brain of the school. Policies, room layouts, class schedules, group records, agent contacts, pricing, Erasmus programme details — everything Stella knows, Iris inherits.
An Iris instance for an accommodation manager inherits different context from one for a Director of Studies. Stella's RAG is role-aware, not flat.
From go-live, Iris watches the live inbox continuously. Every approved draft she sends without edit reinforces her model of that person. Every edit teaches her something new — a preferred phrase, a corrected tone, a relationship nuance she had missed.
A memory compression loop runs every 20 minutes server-side — compressing new signals from inbox, calendar, and CRM into the per-user knowledge base so context stays current without growing unbounded.
The goal: over weeks, drafts require fewer edits. Over months, the approval tap becomes genuinely effortless for known contact types. Trust is earned, not assumed.
The approval loop is the core UX of Iris — and it is non-negotiable at launch. Every draft, every action, every send requires explicit human sign-off. This is not a limitation — it is the feature that makes delegated authority trustworthy.
The approval channel — a private Zulip thread or Telegram chat — is the manager's single pane of glass. Every Iris action arrives there: the original email, the draft reply, the research context. One tap to send. One tap to edit. One tap to skip.
Edits are not wasted. Every correction flows back to Iris as a learning signal. The more consistently a manager approves without editing, the more accurately Iris reflects them.
Over time, as trust is established for specific contact types or message categories, certain actions can be promoted to auto-send — for example, FYI-only acknowledgement emails after a month of zero corrections. Each promotion is explicit and reversible.
Every draft requires explicit approval. Nothing is sent automatically. Full human control.
After 30 days of zero corrections on FYI-only replies, that category may be promoted to auto-send. Explicit opt-in per manager.
Specific recurring contacts — where the pattern is well-established and corrections have been zero — may be trusted fully. Roadmap feature.
Every Iris instance is fully isolated. No shared model. No shared memory. No shared namespace. Each manager has their own private AI — running inside the school's Stella deployment.
Each Iris instance writes to its own isolated vector namespace in the school's Qdrant instance. No manager's sent folder, relationship map, or correction history is visible to any other Iris instance.
iris:{user_id}:sent · iris:{user_id}:memoryCorrection history, approval patterns, contact relationship data, and trust level configuration are stored in a per-user SQLite database. Compact, portable, and fully owned by the school.
iris_{user_id}.dbAlongside the personal namespace, every Iris instance can query the school's shared Stella RAG — for organisational context, policies, schedules, and group records — without cross-contaminating personal data.
stella:{school_id}:sharedIris connects via IMAP to monitor the inbox. This can be a dedicated PA address or the manager's own account. Either way, the connection is read-write (for draft saving) but sends only happen through the approval loop.
IMAP · SMTP · OAuth2 where supportedEvery 20 minutes, a server-side process compresses recent inbox, calendar, and CRM signals into a rolling summary stored in the personal namespace. Context stays current; storage stays bounded.
20-min cycle · LLM summarisation · Qdrant updateDrafts and notifications flow through Zulip (for schools on the Stella platform) or Telegram. The approval channel is per-user, private, and the exclusive communication path between Iris and her manager.
Zulip DM · Telegram · ExtensibleEvery other Stella module operates from a fixed policy or shared knowledge base. Iris is different.
Rachel knows how to assess a student. Sarah knows how to explain grammar. CLARA knows how to plan a lesson. Leo knows how to run a campaign. All of them operate from a shared understanding of the school — not from a personal understanding of an individual.
Iris learns a specific person — their voice, their relationships, their patterns, their preferences. The per-user namespace, the sent folder ingestion, the correction-based learning: these are architectural decisions that exist precisely because no two managers communicate the same way.
That is also why the trust model matters. Iris operates with delegated sending authority — the first Stella module to do so. That requires a more careful, more explicit, more auditable approval architecture than any other module in the suite.
Open 40 emails, decide what needs a reply, draft each one, send
Open your approval channel, tap ✅ eight times, done
Look up group arrival details before replying to a coordinator
Iris already pulled the group record and wrote it into the draft
Someone emails on a Saturday — no reply until Monday morning
Draft ready in your channel within minutes — send when you're ready
Remember to follow up on four outstanding threads from last week
Iris flagged them in Tuesday's briefing with follow-up drafts queued
Write a welcome meeting coordination email across six recipients
"Organise a welcome meeting for the Bologna group" — drafts arrive in 30 seconds
Iris handles the most sensitive data in your school — personal emails, relationship context, individual communication patterns. That is why the architecture is as strict as it is.
Each manager's Iris instance writes to an isolated Qdrant namespace. No data leaks between users. No shared context. Complete separation at the vector store level.
All AI processing happens on the school's own hardware. Sent folder embeddings, inbox classification, draft generation — none of it touches an external AI provider.
Every draft, every approval, every send, every correction is logged. You always know what Iris sent, when, and whether it was edited. Permanent, local, auditable.
Email data is processed locally. No personal communication leaves the school's infrastructure. Consent and data access rights are manageable at the per-user level.
Iris's approval channel means the manager always sees every action before it happens. There are no background sends, no hidden decisions, no invisible automation.
The per-user namespace and SQLite database can be fully deleted at any time — wiping all learned context while leaving other Iris instances and Stella unaffected.
Every Iris deployment runs on your school's own NVIDIA hardware. Sent folder data, inbox content, personal communication patterns, and relationship context are processed on-site and stored locally — never transmitted to external AI providers, never shared between schools, never used to train any model outside your infrastructure.
Iris is a Stella module available to any Buongiorno deployment. Get in touch to discuss an implementation at your school.