dailies.

Inspiration

Release data tells me what happened. It does not always tell me why.

I built Dailies around a simple problem: when a movie's performance changes unexpectedly, someone still has to dig through the data, form hypotheses, test them, and figure out what is actually happening.

I wanted to turn that investigation into an agent.

Instead of giving an AI a dashboard and asking it to summarize charts, I built Dailies to investigate the data, challenge its own assumptions, and produce an evidence-backed explanation.

What it does

Dailies is an AI-powered release intelligence agent that monitors movie performance and investigates unusual changes across markets and metrics.

It can detect anomalies, generate possible explanations, query the underlying data, test those hypotheses, and turn the results into a concise release brief.

For example, Dailies can detect a decline in completion in LATAM. Instead of immediately assuming the cause, it investigates the available data and verifies whether the hypothesis is actually supported.

It can also surface secondary signals, such as an unusual increase in viewing volume in APAC.

The goal is simple:

Don't just tell me that something changed. Investigate why.

How I built it

I built Dailies around an AI investigation agent and structured release data.

I use ClickHouse for analytical queries and Gemini for reasoning and investigation. The agent follows an investigation loop:

[ \text{Detect} \rightarrow \text{Hypothesize} \rightarrow \text{Query} \rightarrow \text{Verify} \rightarrow \text{Explain} ]

The data layer gives the agent access to the underlying metrics, while the reasoning layer decides what to investigate next.

This allows Dailies to move beyond static dashboards and perform an actual investigation.

Challenges I ran into

The hardest part was making the agent investigate instead of simply generating a convincing explanation.

An AI can produce a plausible story very quickly. That does not mean the story is supported by the data.

I had to structure the workflow so hypotheses were treated as things to test, not facts to repeat.

Another challenge was latency. Some investigations require multiple Gemini calls and database queries, which makes a real-time demo slower than a normal dashboard.

I had to design the experience around that reality instead of hiding it.

Accomplishments that I'm proud of

I'm proud that I built a working investigation loop that connects AI reasoning with analytical queries.

More importantly, Dailies can demonstrate a complete investigation from anomaly detection to verified finding.

The LATAM completion decline is a good example. Dailies does not stop at saying, "completion is down." It investigates possible causes and determines what the data actually supports.

That distinction is the core of the product.

What I learned

I learned that building an AI analyst is less about making the model generate better prose and more about giving it the right tools and constraints.

The database is the source of truth.

The model is the investigator.

That separation makes the system much more reliable.

I also learned that good agent design requires explicit investigation states, clear tool access, and a way to distinguish between a hypothesis and a verified finding.

What's next for Dailies

My next step is to make Dailies continuously monitor releases instead of waiting for someone to start an investigation.

I want it to detect meaningful changes automatically, prioritize the incidents that matter, investigate them in the background, and notify teams when there is a finding worth acting on.

Longer term, I see Dailies becoming a release intelligence layer that sits above existing analytics infrastructure.

The dashboard shows the numbers.

Dailies investigates the story behind them.