Maya
At a glance
- Live at: BDO Unibank (card delivery monitoring)
- Channels: background monitoring; escalates via voice (calls Vera) and tickets (D-365)
- First deployed: late 2025
- Owner: Control tower pod
The problem it solves
A card delivery has dozens of touchpoints — pickup, in-transit scans, attempt 1, attempt 2, RTS, redelivery. Anomalies that humans should act on (repeated failed attempts, mismatched recipient, courier silence) get lost in the noise unless someone is actively watching every order.
Maya sits on top of the live order stream, watches all of it, and acts when patterns emerge. Think of it as the difference between a smoke alarm in every room and one person walking through occasionally sniffing.
How it works
What Maya watches
| Signal | Threshold | Action |
|---|---|---|
| No carrier status update | > 12 hours | Auto-alert ops |
| Failed delivery attempts | ≥ 3 in 48 hours | Trigger Vera to re-verify with customer |
| Repeated delivery to wrong location | ≥ 2 instances | Trigger address re-check |
| Recipient name mismatch | Any | Route to fraud team |
| TAT breach imminent | < 6 hours | Notify carrier + customer |
| TAT breached | Any | Auto-reallocate to fallback carrier where eligible |
These thresholds are configurable per customer.
Tech stack
| Layer | What’s used |
|---|---|
| Models | Pattern recognition (rule-based + lightweight classification); narrative summary via LLM when escalating. |
| Memory | Long-term: prior order patterns per customer, per carrier, per route. |
| Tools (MCP) | EDMS, carrier APIs, CCM, D-365, Vera trigger, fraud-team routing. |
| Guardrails | Confidence threshold per action; high-stakes actions require human ack within X minutes. |
| Evals | Synthetic anomaly injection; replay of historical incidents to check detection lag. |
Eval metrics (current)
Placeholder: Need numbers. Suggested: anomalies caught vs. anomalies escaped (precision/recall), mean time-to-detect, false-positive rate, $ savings from reallocations.
Lessons learned
Placeholder: Especially around (a) tuning thresholds without alert fatigue, and (b) how Maya’s actions interacted with humans who were already watching the same order.
Sources
- BDO Unibank deck, “Control Tower Agent ‘Maya’” section
- Maya video walkthrough (Atlas, same architecture): ATLAS YouTube link
- See BDO case study for context
- Related: Atlas (broader version of the same idea)
Changelog
- 26 May 2026: Initial draft from BDO deck.