Andromeda
Andromeda is KlayrAI’s user intelligence layer. It builds a behavioral profile of each user over time and uses that profile to adapt every diagnostic, recommendation, and report to match your decision-making style.Why personalization matters
Two media buyers looking at the same campaign data may need completely different advice:- A conservative buyer managing a client’s budget wants cautious, low-risk suggestions with proven track records
- An aggressive growth marketer spending their own money wants bold scaling recommendations that maximize upside
The UserBehaviorProfile
Every KlayrAI user has aUserBehaviorProfile that captures their preferences and patterns:
| Field | Type | Description |
|---|---|---|
riskAppetite | LOW, MEDIUM, HIGH | How much risk the user is comfortable with in budget changes and targeting experiments |
focusMetric | ROAS, CPA, CPM, CTR, VOLUME, REACH | The primary KPI the user optimizes for |
preferredActions | string[] | Actions the user has historically approved (e.g., “increase_budget”, “refresh_creative”, “narrow_audience”) |
avgManagedBudget | number | Average daily budget across all managed campaigns |
recommendationApplyRate | number | Percentage of KlayrAI recommendations the user has approved and applied |
How the profile is built
Andromeda computes the profile from theUserEvent table, which tracks every meaningful interaction:
Events tracked
| Event | Signal |
|---|---|
| Recommendation approved | Indicates comfort with that action type and risk level |
| Recommendation dismissed | Indicates discomfort or disagreement with the suggestion |
| Manual campaign changes | Reveals preferred optimization patterns |
| Diagnostic frequency | Shows how actively the user monitors campaigns |
| Dashboard focus | Which metrics and views the user spends time on |
| Budget change patterns | Historical budget adjustments reveal risk tolerance |
Computation schedule
Profiles are recomputed weekly via a scheduled job. The computation looks at the trailing 30 days of user events with a decay factor — recent behavior is weighted more heavily than older behavior.New users start with a default profile (MEDIUM risk appetite, ROAS focus metric) until enough behavioral data is collected — typically after 5-10 diagnostic interactions.
How Andromeda adapts diagnostics
The user’s Andromeda profile is injected into every Claude diagnostic call as context. This changes the analysis in several ways:Risk appetite adaptation
LOW risk appetite
LOW risk appetite
- Recommendations are conservative and incremental
- Budget changes are capped at 10-15% adjustments
- “Wait and monitor” is suggested when data is ambiguous
- Emphasis on protecting current performance
- Language: “Consider”, “You might want to”, “A safe approach would be”
MEDIUM risk appetite
MEDIUM risk appetite
- Balanced recommendations with moderate ambition
- Budget changes of 15-30% suggested when evidence supports it
- Both upside potential and downside risks are presented
- Language: “We recommend”, “The data supports”, “This should improve”
HIGH risk appetite
HIGH risk appetite
- Aggressive scaling and experimentation encouraged
- Budget changes of 30-50%+ when opportunity exists
- Focus on maximum upside with acceptable risk
- Faster iteration cycles recommended
- Language: “You should”, “The opportunity is clear”, “Act now to capture”
Focus metric adaptation
The user’s primary KPI determines how issues are prioritized and which recommendations come first:| Focus metric | Prioritization behavior |
|---|---|
| ROAS | Efficiency issues ranked highest. CPA creep and budget waste flagged aggressively. |
| CPA | Cost-per-acquisition is the north star. Learning phase and creative fatigue prioritized. |
| CPM | Auction costs monitored closely. Overlap and audience saturation ranked highest. |
| CTR | Creative performance emphasized. Fatigue and ad relevance issues prioritized. |
| VOLUME | Conversion volume is key. Scaling opportunities and budget headroom highlighted. |
| REACH | Audience expansion opportunities prioritized. Saturation flagged early. |
Recommendation filtering
Andromeda tracks which types of recommendations users historically approve. If a user consistently dismisses “expand audience” suggestions but approves “refresh creative” suggestions, future diagnostics will:- Rank creative-related recommendations higher
- Provide more detail and confidence data for audience recommendations (to build trust)
- Frame audience suggestions in terms the user responds to
Example: Same campaign, different profiles
Consider a campaign with declining ROAS (2.1x, down from 3.0x) and rising CPA.Profile A: Conservative, ROAS-focused
Risk level: HIGH Your ROAS has declined 30% over the past 7 days. The primary driver is creative fatigue in your “Broad Lookalike” ad set. Recommendation: Pause the underperforming creatives and introduce 3 new variations based on your top-performing angles. This is a low-risk change that should stabilize ROAS within 5-7 days. Consider reducing the daily budget by 15% until performance recovers.
Profile B: Aggressive, VOLUME-focused
Risk level: MEDIUM Your campaign is showing signs of creative fatigue, but overall conversion volume remains strong at 142 this week. ROAS has dipped to 2.1x. Recommendation: Keep current spend levels to maintain volume. Immediately launch 5 new creatives to rotate against fatigued ones. If new creatives perform, increase budget by 30% to capitalize on the refreshed creative set. The volume opportunity outweighs the temporary ROAS dip.
Same data, same issues — completely different tone, recommendations, and priorities. This is the value of Andromeda.
API access
When using the API, theandromedaContext field in diagnostic responses shows how the profile influenced the analysis:
Privacy
- Andromeda profiles are stored per-user and scoped to their workspace
- Profile data is never shared between users or workspaces
- Users can reset their profile at any time from Settings > Andromeda
- Profile data is not used to train AI models