e2emasa

Growth experimentation, Adaptive systems

TITAN
Marketing Agent

A hypothesis earns its next step.

A marketing experimentation engine that proposes ideas, checks drafts against an author’s voice, measures outcomes and carries the evidence into its next decision.

An experiment, then evidence.

  1. HypothesesProject brief and learnings
  2. Thompson plannerChannel × category
  3. Draft & voice gateGenerate, critique, review
  4. Approval policyFull, approval, suggest only
  5. Channel actionPublish or manual handover
  6. Measured outcomesClicks, conversions, engagement
  7. Deterministic rewardWeighted metrics, normalized
  8. LearningsQualitative insight memory
Measured results feed both channel selection and future hypotheses. A per-channel policy controls whether a draft may be published or requires review.
  • Hypotheses connects to Thompson planner.
  • Thompson planner connects to Draft & voice gate.
  • Draft & voice gate connects to Approval policy.
  • Approval policy connects to Channel action.
  • Channel action connects to Measured outcomes.
  • Measured outcomes connects to Deterministic reward.
  • Deterministic reward connects to Learnings.
  • Learnings connects to Hypotheses.
  • Deterministic reward connects to Thompson planner.

A hypothesis earns its next step

TITAN Marketing Agent connects strategy, writing and measured results in one cycle. It maintains a hypothesis backlog, selects a channel and category through Thompson sampling, and develops drafts against the project brief and voice profile. Independent critique, editorial review and deterministic phrase checks decide what reaches the approval queue.

Each channel has its own autonomy and cadence policy. Approved work is published or handed over for manual posting; tracked clicks, conversions and platform engagement form the evidence. Code computes the reward, while the model develops qualitative learnings. The CLI and dashboard share the same experiment state transitions, and completed results inform future selection.

01

Write in a real voice

Approved writing samples and learned rejection notes guide new drafts. Best-of-N generation, a separate critic and editorial review work alongside deterministic validation.

02

Measure what happened

Tracked links and experiment-scoped metrics connect actions to outcomes. Retries and cumulative platform measurements are handled without inflating the result.

03

Learn without drifting

The planner rebuilds arm statistics from completed results. Recent evidence carries more weight; approval, retry and manual-posting rules are shared by the dashboard and CLI.

Keep evidence current

A result loses half its selection weight after each configured half-life. Recent outcomes guide the next experiment, while older evidence gradually carries less influence.

Selection weight of a completed result
Now1.001 half-life0.502 half-lives0.25

weight = 0.5 ^ (age / half-life)

Explore the engineering

Technical documentation and a runnable reference core.

GITHUB