Rebuilding an AI-Generated App for Stability and Scale (feat. Chris Pearcey)
AI-powered “vibe coding” accelerates MVP development, but sustainable scale requires human-led architecture, clean data modeling, and refactoring.
In this conversation, Chris Pearcey, founder of Decisio, explains that his mission is to restore decision-making power to users overwhelmed by choice. In an environment where recommendation engines increasingly optimize for platform revenue rather than user clarity, Decisio attempts to formalize explicit preference capture through structured interaction.
“Right now in the world, decision is a difficult thing because there are way too many choices.”
This framing establishes the technical challenge: discovery systems today infer preference indirectly. Decisio aims to replace inference-heavy engagement modeling with explicit user intent signals.
Podcast
Rebuilding a Consumer App for Stability, Performance, and Real Traction — on Apple and Spotify.
Explicit Intent vs Algorithmic Steering
Traditional streaming platforms rely on behavioral inference. They track watch duration, replays, pauses, and abandonment patterns to model preference. These systems optimize for engagement and monetization alignment rather than authentic user satisfaction. Decisio introduces a four-direction swipe system that captures seen/like, seen/dislike, want-to-watch, and not-interested. This reduces ambiguity and increases signal quality for downstream recommendation modeling.
“We’re trying to give folks the ability to use both their intent on items and their taste on items to swipe through existing content to learn what we actually like.”
From a data architecture perspective, this changes everything. Instead of inferring preference from noisy behavior, the system stores discrete, user-declared events. That produces cleaner relational modeling opportunities and simplifies cohort segmentation.
Vibe Coding the MVP, is Phase 1!
The initial goal was speed, not perfection. Pearcey drafted the business plan with AI assistance, used Copilot to structure requirements, and leveraged AI-enabled Figma tooling for rapid UI prototyping. The first swiping implementation was functional within roughly ten days.
“I developed the existing app… I did 100% of that myself with vibe coding initially.”
Approximately 95% of the first frontend iteration was AI-generated, with about 70% of backend functions scaffolded by AI before manual cleanup. This approach dramatically reduced time-to-beta and validated demand before capital was raised.
Vibe coding served its purpose: fast market validation. But it was never intended to be the long-term engineering foundation.
Where AI Accelerated Development
AI proved extremely effective in specific domains. It handled UI scaffolding, component generation, boilerplate backend functions, business plan formatting, and marketing creative production. It even functioned as a lightweight program manager in early discussions.
“Copilot actually helped as a program manager.”
However, as complexity increased, limitations emerged. AI occasionally introduced inefficient logic, redundant code, and fragile dependencies. Performance tuning required human review. In some cases, manual refactoring reduced response times from seconds to milliseconds after identifying inefficient query paths.
Pearcey summarized the balance clearly:
“AI is a great companion. It’s not a great lead engineer yet.”
This reflects a broader truth: generative tools are strong at scaffolding but still require experienced engineering judgment for optimization and architectural consistency.
Architecture and Stack Decisions
Decisio operates primarily on Google Cloud. The operational database runs on Cloud SQL Postgres, with BigQuery handling analytics. Firebase hosts the frontend, and the backend runs on Node.js with React-based clients. The application currently ships as a Progressive Web App wrapped for both iOS and Android, enabling rapid release cycles without extended App Store approval delays.
The decision to use Postgres over NoSQL was founder-driven rather than AI-suggested. Explicit swipe data creates structured relational records involving users, titles, providers, genres, and preference states. A normalized relational schema simplifies aggregation, cohort matching, and analytics joins.
Separating transactional storage (Postgres) from analytical processing (BigQuery) creates architectural clarity early, reducing future migration complexity.
Data Integrity and the Limits of AI
One major turning point involved content attribution. AI-generated mappings between titles and streaming providers proved unreliable, producing roughly 40% false positives. Given that accurate content availability is foundational to user trust, this was unacceptable.
The team shifted to a traditional ETL pipeline approach, purchasing access to a third-party content API and implementing nightly delta syncs into BigQuery.
“I actually had to build an old school data pipeline ETL… and honestly, it’s been the best decision I made.”
This highlights a critical lesson for AI-era startups: generative systems are excellent for scaffolding, but deterministic pipelines are essential for correctness.
Testing in Early-Stage Consumer Systems
Testing remains pragmatic rather than enterprise-level. The team performs manual regression tests across multiple Android devices and iOS hardware, validating core user flows before release.
“We are really just testing… opening our phones and going through our test cases.”
Automated testing coverage is limited but acknowledged as the next maturation step. This reflects a common startup tradeoff: speed and feedback cycles outweigh full automation early on, but infrastructure hardening must follow validation.
Refactoring for Scale
Once traction was validated, the system entered a deliberate rebuild phase. Dedicated engineers began restructuring backend functions, improving environment separation, and strengthening DevOps processes.
“They’re rebuilding everything now. They’re resetting the foundation.”
This transition—from AI-heavy prototype to structured engineering rebuild—is often where startups either succeed or collapse. Addressing technical debt before user growth stresses infrastructure is essential.
AI and Team Size Compression
AI did not eliminate engineering roles, but it reduced team size requirements. Pearcey estimates the company will likely remain between 15 and 20 employees at scale—far smaller than similar startups from a decade ago.
“We’ll never probably be more than 15 to 20 employees ever because of AI.”
AI reduces boilerplate coding, accelerates marketing asset generation, and compresses DevOps overhead. However, it does not replace strategic thinking, data modeling expertise, domain understanding, or sales execution. The shift is not toward zero engineers but toward higher-leverage engineers.
Progressive Web Apps and Native Migration
The current PWA approach enables rapid iteration and immediate deployment without multi-day App Store approval cycles.
“We can… make multiple changes in a single day if we need to.”
However, the long-term roadmap includes migrating to native apps for richer push notifications and deeper system integration once the feature surface stabilizes.
Speed first. Polish later.
Building Domain-Specific Intelligence
Looking forward, Decisio aims to leverage its structured swipe dataset to create domain-specific intelligence models. The long-term vision includes building retrieval-augmented generation systems for entertainment discovery.
“I want to be the source of truth for LLMs for entertainment.”
By owning high-quality, explicit intent data, the company could build a differentiated intelligence layer that general-purpose LLMs cannot easily replicate.
Closing Reflection: Engineers with AI
Pearcey’s journey illustrates a nuanced reality. AI dramatically compresses early-stage build timelines and reduces staffing requirements. It accelerates prototypes and lowers barriers to entry. But it does not eliminate the need for architectural thinking, data modeling discipline, performance optimization, or human review.
The old startup model required large teams and long pre-launch cycles. The new model enables rapid AI-assisted validation followed by deliberate refactoring and targeted hiring.
The barrier to entry is lower. The barrier to building durable systems is not. The future is not AI replacing engineers.
It is engineers who understand architecture, data integrity, and system design—leveraging AI as a multiplier rather than a substitute.