Inspiration

What if the loudest moment in a stream is not the best guide to what the creator should make next?

Imagine a musician trying original songs during a Twitch stream. A viewer gifts subscriptions, chat fills with excitement, and the segment looks successful. The musician now faces a real decision: should they spend their next evening preparing a full original-song show?

The excitement is real—but the reason behind it is not immediately clear. Did viewers love the format, were they celebrating the gifts, or did both things strengthen each other?

That question inspired StreamPulse. I wanted to help creators investigate a promising idea before committing hours of preparation to it.

The goal is not to dismiss gifts or label community enthusiasm as fake. Community support can be part of what makes a format successful. The goal is to help creators understand whether a result is repeatable and whether repeating it is worth the effort.

What it does

StreamPulse helps Twitch creators answer one practical question: “Is this format worth spending more time on?”

Instead of stopping at a dashboard, it follows a decision from the creator’s initial question through a comparison, a recorded choice, and a later follow-up.

The creator chooses two formats, defines what success would mean, and records their preparation budget. During streams, they mark format changes through simple controls or an OBS dock. StreamPulse brings together available audience observations, chat activity, gifting events, preparation effort, and collection gaps so the creator can review what happened.

Its private conversational assistant makes that workflow accessible without requiring the creator to navigate every screen. They can describe a comparison, get guided through missing details, operate supported live controls, review saved records, and record their decision. Changes go through the application’s permissions and approval process, with a saved receipt showing what actually happened.

Optional audience feedback lets viewers contribute without paying or receiving extra voting weight for gifting. Those responses remain feedback from participating viewers—not a claim about the entire audience.

The outcome is not always “Format A wins.” Sometimes the useful answer is “Compare again before increasing preparation” or “There is not enough evidence yet.”

How we built it

I used OpenAI Codex throughout development to turn the project design into the interface, backend services, Twitch integration, conversational assistant, analysis engine, and test infrastructure. The work progressed through implementation, testing, browser checks, and revisions rather than stopping at generated screens.

The application uses React and TypeScript for the creator workspace, Fastify for the API, PostgreSQL for persistent records, and Python for numerical analysis. Background workers handle Twitch collection and processing independently of the creator’s browser. Twitch OAuth, EventSub, and API integration connect the application to the broadcaster and dedicated bot account.

A key architectural choice was separating conversation from calculation. The assistant helps interpret requests and guide the workflow, while ordinary application code controls permissions, executes actions, computes results, and records outcomes. Connecting a model does not automatically authorize it to process audience data; approved context is controlled separately.

I also built a separate synthetic demonstration environment so the workflow can be explored without Twitch credentials or changes to a connected creator workspace. Automated tests, runtime checks, and container-based checks support the implementation.

Challenges we ran into

Avoiding a misleading answer. The hardest conceptual challenge was resisting the tempting shortcut of subtracting activity around gifts and calling the remainder “real engagement.” A compelling performance might itself inspire gifting. StreamPulse therefore preserves gifts as context and uses defined comparisons rather than claiming it can perfectly separate the causes of excitement.

Making the assistant safe to act. A conversational interface becomes much more demanding when it can change records instead of only answering questions. I had to handle approval previews, stale information, interrupted requests, retries, and concurrent actions. The implementation preserves pending work and checks authoritative results before showing an action as completed.

Connecting a real Twitch workflow. Account authorization, bot identity, webhook delivery, stream state, and audience sampling all needed to work together. During setup, the account guard rejected an authorization using the broadcaster where the dedicated bot was expected. Controlled broadcasts then helped verify the integration beyond simulated events.

Keeping uncertainty understandable. Missing data, interrupted streams, and small comparisons should not turn into confident recommendations. The interface needed to explain those limitations without overwhelming a creator who simply wants to plan their next show. That led to short recommendations with supporting evidence available underneath.

Accomplishments that we're proud of

I am proud of building a connected workflow rather than an isolated analytics screen: plan → stream → review → decide → follow up.

The project includes a working private assistant, guided actions, live format controls, an OBS dock, evidence review, preparation tracking, and a history that keeps the application’s recommendation separate from the creator’s actual choice.

Controlled Twitch broadcasts verified stream online/offline events, real chat collection, format controls, and viewer sampling. A later rehearsal also recorded chat activity within the comparison window and a feedback response. Local container checks and GitHub Actions exercised the application stack, including its Python analysis service. These tests establish technical progress—not proof that the product already improves creators’ income or content decisions.

What we learned

The biggest lesson was that more activity is not automatically better evidence. A dramatic spike can be interesting without being enough to justify a larger creative commitment.

I also learned that useful AI assistance depends on the system around the model: clear permissions, saved state, reliable actions, understandable confirmations, and recoverable failures.

Finally, I learned to separate three questions: Does the code work? Does the integration work in a real stream? Does the product genuinely help a creator? Each needs its own kind of validation.

What's next for StreamPulse

The next step is testing with creators making real programming decisions over several streams. I want to measure whether StreamPulse improves their choices, reduces avoidable preparation, and fits naturally into their routine.

Further work includes live verification of gift, raid, and shared-chat scenarios, broader accessibility testing, and production deployment and recovery checks. Those remain distinct from the synthetic tests and controlled broadcasts already completed.

Built With

Share this project:

Updates

Submission history