AI Growth Doctor

AI Growth Doctor turns growth metrics into auditable daily decisions, combining specialist agents, deterministic guardrails, calibrated forecasts, and human judgment

Comment

Inspiration

I am a solo founder of Hitung Kalori, a calorie tracking app in Indonesia. Almost every morning I need to answer one simple but difficult question: should I scale growth today, or should I fix something first?

The problem is the answer is never from one metric only. Activation can look okay, but retention can be weak. Ads can look cheaper, but users may not stay. Monetization can show some signal, but the sample can be too small. A normal dashboard can show all the numbers, but it does not help me decide what to do.

AI Growth Doctor comes from this real daily problem. I wanted to build a system that can read real product data, let several AI specialists look at it from different angles, challenge each other with evidence, and then give me a safer daily operating decision.

What it does

AI Growth Doctor is a Qwen-powered multi-agent decision system for mobile app growth.

Every run starts from real metrics, then the system runs six specialist agents in parallel:

  • Activation Agent
  • Retention Agent
  • Monetization Agent
  • Version Agent
  • Ads Agent
  • Tomorrow Forecast Agent

Each agent only focuses on its own area first. After that, the system runs a structured negotiation step. In this step, agents can raise objections, warnings, constraints, or partial concessions. They cannot just argue freely. They need to refer back to the metrics.

After the negotiation, the final decision agent creates a daily operating verdict, for example HOLD_AND_OPTIMIZE. It also gives action items, evidence, risk controls, and metrics to evaluate the result later.

In one real run, the system found that users who reach the workspace are actually good users and most of them can log food successfully. But around half of the sessions drop before reaching the workspace. Retention also drops a lot from D0 to D1. Because of that, the ads agent did not recommend aggressive scaling. The final decision was to fix onboarding and D1 habit first, before increasing ad spend or paywall pressure.

How we built it

AI Growth Doctor is built with Laravel. It has a backend pipeline, dashboard UI, graph view, interaction log, audit trace viewer, forecast evaluation, and Qwen-powered agent execution.

The pipeline is like this:

  1. Load daily growth checkpoint from real app data.
  2. Extract deterministic metrics for activation, retention, monetization, ads, and app versions.
  3. Apply guardrail policy and metric mapping validation.
  4. Run six Qwen specialist agents in parallel using Alibaba Cloud Qwen / DashScope OpenAI-compatible API.
  5. Run structured negotiation using evidence-bound claims.
  6. Build compact context for the final decision agent.
  7. Run the final decision agent.
  8. Run a scenario simulator to compare the recommended action with a baseline.
  9. Store the full audit trace so every run can be checked again.

The system is deployed on Alibaba Cloud ECS and uses Qwen through the DashScope compatible endpoint. The dashboard shows the graph flow, interaction log, forecast evaluation, calibration score, baseline comparison, and full run trace.

Challenges we ran into

The biggest challenge was making the agents collaborate in a useful way, not just create a normal chatbot conversation.

At first, multi-agent debate can easily become too free-form. It can look smart, but the conflict may not be real. So I made the negotiation more structured. Every objection or warning must have evidence references. The system classifies each point as support, warning, constraint, concession, or material conflict.

Another challenge was context size. Some agents need a lot of data, especially activation, monetization, and version agents. But the final decision agent should not receive too much unnecessary detail. So I built compact context builders to keep the final decision focused and reliable.

Forecasting was also not simple. I did not want the system to blindly trust its own forecast. So I added forecast evaluation and calibration. If forecast accuracy is not strong enough, the forecast becomes only a supporting guardrail, not the main decision owner.

The last challenge was using real business data while still making the system easy to demo. That is why every run has a run ID, timestamp, graph view, interaction log, and full audit trace.

Accomplishments that we're proud of

I am proud that AI Growth Doctor is not only a dashboard and not only a single AI summary.

It shows a real Agent Society workflow:

  • Six Qwen agents run in parallel.
  • Each agent has its own role.
  • Agents can challenge each other with evidence.
  • The system can resolve tensions without forcing fake debate rounds.
  • The final decision gives action items, success metrics, stop-loss metrics, and rollback conditions.
  • The full trace can be audited.

In one real run, the system processed 32 interaction sequences, completed six specialist agents, ran structured negotiation, did forecast self-calibration, generated a final operating decision, and produced a scenario simulation in under two minutes.

I am also proud that the system does not always recommend growth. In this run, it found onboarding and retention problems, blocked aggressive scaling, and recommended guarded optimization instead. For a real business, this is important because bad scaling can waste money fast.

What we learned

I learned that a useful AI system for business should not only give a confident answer. It should also know when the evidence is weak and when the action should be careful.

I also learned that multi-agent systems are much better when the collaboration is structured. Free-form debate is interesting, but evidence-based negotiation is more useful for real decisions.

Another lesson is that forecast should not automatically control the decision. It should earn trust over time. That is why the system tracks forecast hit rate and uses mature actual data to calibrate itself.

Most importantly, I learned that AI can help a solo founder work more like a small growth team. One system can act like an activation analyst, retention analyst, monetization analyst, ads analyst, release analyst, forecaster, and decision assistant working together every day.

What's next for AI Growth Doctor

Next, I want to make AI Growth Doctor more connected and more actionable.

Planned improvements:

  • Direct integrations with GA4, Google Ads, Play Console, and subscription data.
  • Better forecast calibration and confidence scoring.
  • Experiment tracking and post-action evaluation.
  • Daily decision summary through Slack or Discord.
  • Action history, so the system can learn which recommendations worked.
  • Templates for other businesses, not only mobile apps, for example SaaS, e-commerce, and subscription products.
  • Safer execution workflows where approved actions can create tickets, campaigns, or experiment setup.

The long-term vision is to make an AI operating layer for founders. Not just a system that explains what happened, but a system that helps decide what to do next, with evidence, negotiation, and accountability.

Built With

Share this project:

Updates