Dodo Engineering · March 2026
Outreach Time
Intelligence
Adaptive send-time optimization via Thompson Sampling — how Dodo learns when to reach each clinic's clients for maximum engagement.
The Problem
Everyone sends at 9 AM.
Nobody asks if that's the right time.
Attention competition
All recipients are reached simultaneously, competing with every other business using the same strategy.
Missed optimal windows
One clinic's clients may respond best at 11 AM; another's at 3 PM. A fixed schedule never discovers this.
No learning
The system doesn't adapt based on what has worked before. Every day is the same guess.
The Framework
The Multi-Armed Bandit
Imagine a gambler facing a row of slot machines, each with an unknown payout probability. The gambler must balance exploration (trying unfamiliar machines) against exploitation (playing the best-known machine).
In send-time optimization, each arm is a time slot and each pull is sending a message. The outcome is binary:
Engaged
Client responded within 24 hours
Did not engage
No response within the window
The Algorithm
Thompson Sampling
Instead of maintaining a single estimate, Thompson Sampling keeps a full probability distribution for each time slot. Uncertain slots have wide distributions (more exploration). Well-understood slots have narrow ones (more exploitation).
0 observations
Beta(1, 1) — Complete uncertainty
50 observations
Beta(11, 41) — Emerging signal
200 observations
Beta(41, 161) — High confidence
The update rule is remarkably simple: client engaged → increment α. Client didn't engage → increment β. No matrix inversions. No gradient descent. No retraining. Just two integer increments per observation.
Applying It
11 hourly arms,
one optimal window.
The outreach window spans 9 AM – 8 PM local time, discretized into 11 one-hour slots. Each arm learns independently per business, with actual send times randomized within the winning slot.
Convergence
How fast does it learn?
Exploring
Posteriors still wide; heavy exploration across all slots.
Converging
Best arm identified. Arms differing by ≥5pp clearly separated.
Exploiting
Near full exploitation. Fine-grained differences detectable.
Bootstrapping
New businesses don't start from scratch.
Global priors computed from aggregate historical data give every new business a head start. Using relative-lift normalization, we extract time-of-day patterns that are consistent across campaign types — even when absolute engagement rates vary wildly.
Without global priors
- Start
- Beta(1, 1) — uniform
- Convergence
- ~6 weeks
- Early sends
- Pure exploration
With global priors
- Start
- Beta(α₀, β₀) — informed
- Convergence
- ~2–3 weeks
- Early sends
- Guided exploitation
The system at a glance
- Algorithm
- Thompson Sampling with Beta posteriors
- Arms
- 11 hourly slots, 9 AM – 8 PM local time
- Scope
- Per-business, shared across campaigns
- Engagement
- Binary; response within 24h (configurable)
- Storage
- 22 integers per business (2 per arm)
- Convergence
- ~2,200 sends; faster with global priors