StockCast

Early in the operation, the volume of installs grew consistently. Product instrumentation, however, revealed an invisible bottleneck: very few users reached the system's main action (playing a conference). Acquisition was optimized for install volume, bringing in users with no alignment to the value proposition.

We changed the strategy to optimize campaigns for qualified in-app events. That is when the analytics architecture failed. We tried to maintain architectural purity by using a single tool (PostHog) to guarantee privacy and data control. But when integrating those mobile events with Google Ads, the abstraction broke: the ad platform read the app events as web traffic, invalidating mobile optimization.

The decision: We abandoned the purity of a single tool and integrated the Google Analytics SDK directly into the app.

The result: With acquisition corrected and new notification journeys introduced, the activation rate (defined as users who open the app and play an audio within equivalent 30-day windows) jumped from 6% to 26%.

The JOBE principle: Abstractions should simplify the system. When architectural purity creates more operational complexity than a direct integration, pragmatism must win. Instrumenting the real funnel matters more than vanity metrics.

Indexing financial documents and conferences from more than 300 companies requires monitoring public sources (CVM/B3 and investor-relations sites) that offer no free push mechanisms. The alternative is periodic polling, which creates a critical infrastructure trade-off: latency × cost × risk of IP blocking.

We built a pipeline that operates 24/7, using internal triggers and batch execution to minimize unnecessary requests. The system detects, records, and fires notifications for new documents in approximately 2.5 minutes, on average, with no human intervention in the normal flow.

However, automation runs into the reality of unstructured data. Identifying URLs across hundreds of investor-relations sites with distinct HTML structures still requires maintenance. And in processing financial analyses derived from those documents, the risk of error (hallucination) is unacceptable.

The decision: We exhaustively automated the ingestion, media re-encoding, and notification infrastructure. But we kept human-in-the-loop and rigorous supervision in the generation and validation of financial analyses.

The JOBE principle: Automation ends where reliability is not yet sufficient. Production systems require knowing what to automate and, above all, where human intervention is non-negotiable to protect the product's value.

Generating deep analyses and podcasts from long audio files and financial PDFs exposes the limitations of AI when it is treated as a "black box". Trying to generate the final content directly from raw sources results in inconsistencies, loss of context, and high computing cost.

The architectural decision: We inserted an intermediate artifact. The system first consolidates a structured, segmented analysis. Only from that validated artifact do we generate the final products (podcasts and text summaries). This enables reuse, consistency, and empirical validation.

In addition, model selection is treated as a continuous engineering decision, not a fixed bet. In transcription tests, we found that generalist models outperformed models dedicated exclusively to audio. The hypothesis? In domains dense with acronyms, indicators, and financial jargon, broad contextual knowledge resolves ambiguities that the acoustic signal alone cannot capture.

The JOBE principle: AI in production is not about using the model of the moment. It is about decomposing workflows, creating validatable intermediate artifacts, swapping models according to the cost/quality relationship, and understanding the business domain to evaluate results.

StockCast has not yet reached commercial product-market fit, and its monetization is still in a discovery phase. But from an engineering standpoint, it fulfills its role: proving that problems of scale, observability, integration, and automation are solved with precise diagnosis, not with far-fetched rewrites.

When your technical foundation starts to limit your business's growth, the next step does not have to be starting from scratch. You need to identify what is actually holding the product back and evolve the existing base with the least possible risk.

JOBE works on exactly this challenge. We strengthen technical foundations so that growing products gain more reliability, observability, and capacity to evolve, turning bottleneck engineering into an engine for your business's expansion.