← Back to the fleet FTE-03 · AI Chief of Staff

Aria

The teammate who never sleeps on your inbox. Two Gmail accounts and WhatsApp watched 24/7, every message LLM-classified, replies drafted in your voice, and delegations that chase themselves until they're done.

Deployed MVP Comms automation Gmail ×2 · WhatsApp · Calendar AI-native 4/5 FTE-fit 4/5

LLM workloads

6Classify · reply · learn…

Urgency model

0–100Mapped to priority

Backend tests

~114Across 11 files

Deploy

Live VMnginx · systemd · watchdog

What runs today vs. the roadmap

Running today

Real LLM, deeply woven — not bolted on.

✦Two inboxes + WhatsApp, watched 24/7 — Gmail polled every 120s with smart deep-sync; WhatsApp every 10s with JID→contact resolution.
✦LLM classification of every message — intent, a 0–100 urgency score, entities, summary and a proposed due date.
✦Replies in your voice — drafts from a per-contact style profile, and a learning loop that improves from every edit you make.
✦Delegations that close themselves — detects a hand-off, opens and assigns a task, chases the overdue via WhatsApp + a gentle email, and auto-closes when the reply says "done."
✦Approve → send, for real — calendar awareness, WebSocket live updates, audit log, and a 13-page React control panel on the live API.

On the roadmap

From smart drafting to true agency.

◆Agentic chat — wire the already-defined tool_calls schema so "type a command, it executes" actually runs (today it's read-only Q&A).
◆Policy-gated autonomy — "auto-send under $X / below urgency Y," trust lists, confidence thresholds, one-click undo. (Autonomy ships in safe testing-mode today.)
◆Semantic memory — replace keyword/ILIKE retrieval with embeddings for sharper voice consistency.
◆Multi-tenant + secrets hygiene — accounts, per-tenant credential vaults; rotate the keys currently in the repo before any public showcase.

Honesty note: six real LLM workloads run, but the actions are still rule-gated, not an agentic planner. Provider is Gemini + OpenAI.

The role it replaces

An EA and an ops coordinator.

Reading, triaging and drafting across email + WhatsApp (the EA), plus opening, assigning and chasing every delegation to closure (the coordinator) — both running on a 24/7 shift.

20–35hrs/week reclaimed · target

The 60-second showcase

veloce://fte-03 — delegation loop

Send the watched inbox: "POS is down at Sehar, please fix ASAP."

Live on screen: classified TOP_URGENT, a delegation task is auto-created and assigned, and the owner gets a WhatsApp alert within ~30 seconds.

Reply "fixed, replaced the router" — and watch the task auto-close from the completion signal.

Aria is the most on-narrative asset in the fleet — a concrete inbox-and-delegation chief-of-staff that already runs around the clock and demos beautifully. The honest gap is agency: the intelligence to read, understand, draft and prioritise is all there; the next release adds the policy-gated autonomy to decide and act end-to-end.