S12 Bonus: Rickard's Deterministic Return: Converting Scattered Data into Autonomous, Production-Grade Apps with Rickard Hansson, Founder & CEO of Gainable

Episode Summary

Executive Summary: Ricard Hansen explains Gainable, an internal-tools platform that shifts AI from direct code generation to data-driven spec creation plus deterministic compilation. The approach aims to solve vibe coding’s fragility, reduce dependence on frontier models, and enable secure deployment behind firewalls for enterprises, with low token costs and repeatable builds.

Main Topics: Weavy update and the bridge to Gainable (Priority: 4/5): Ricard says Weavy has matured upstream into enterprise contracts and big clients, and that the same underlying tech now powers Gainable. Why vibe coding falls short (Priority: 5/5): He argues existing vibe-coding tools break down after a few prompts, overcomplicate simple apps, and are poorly suited to internal tools and non-technical users. Gainable’s data-first workflow (Priority: 5/5): Instead of prompting directly for code, users upload data like spreadsheets; an LLM analyzes the data, suggests app structure, and then a compiler builds the app. Determinism over probabilistic codegen (Priority: 5/5): Ricard emphasizes that the LLM is only used for judgment and spec creation; deterministic validators and a compiler ensure reliable, repeatable output every build. Free-range coding and model-agnostic risk (Priority: 4/5): He criticizes agentic coding systems for choosing overly complex stacks like React/Tailwind simply because they are common in training data, not optimal for the task. Token economics and model dependency (Priority: 4/5): He argues tokens remain heavily subsidized, but that economics will tighten, making model-based products vulnerable if costs rise or access changes. Enterprise deployment and future roadmap (Priority: 5/5): Gainable is aiming for private-cloud, air-gapped, on-prem deployments and is developing its own model on NVIDIA Nemotron to support that enterprise posture.

Key Arguments: Vibe coding is impressive for builders who can articulate intent, but it fails for many users who cannot clearly describe the app they want. Internal tools are a larger and more practical market than consumer SaaS-style apps, but current AI builders are not optimized for that use case. The LLM should be moved earlier in the pipeline to analyze data and define intent, while code generation should be handled by a deterministic compiler. Probabilistic code generation is inherently unreliable; validators and structured contracts are needed to stabilize outputs. Using the simplest or fastest model can work if the system is designed around deterministic checks rather than raw model capability. Model choice is not a durable moat; the real moat is the factory, workflow, brand, and enterprise-ready deployment path. AI token pricing is likely subsidized today, but products built on expensive model usage could be squeezed when pricing normalizes. Enterprise customers need models and deployments they can control locally, including behind firewalls and in air-gapped environments.

Data Points: Build time: 45 to 50 seconds - Ricard says Gainable can generate a full internal app in roughly this time. Application scope: 8-10 views; 12-13 data models - He describes a single build producing a multi-view app with multiple models and CRUD/API endpoints. Backend token cost per app: $1 to $1.50 - He says Gainable’s compiler keeps inference costs low enough to build a complete app cheaply. Weavy focus shift: Enterprise / upstream - He notes Weavy has moved from consumerized competition toward larger enterprise clients and contracts. Model choice example: Haiku 4.5 - He says he challenged himself to build using a “dumbest model” rather than relying on the latest frontier model. Validation layer: Almost 100 deterministic validators - He says Gainable uses around a hundred validators to snap probabilistic outputs into correct structure.

Pivotal Quotes: "I use AI to build a spec that is following a specific format I have." — Ricard Hansen: Explaining that Gainable uses AI for structured planning, not direct code generation. "When you click build, LLM is doing basically nothing." — Ricard Hansen: Describing how Gainable hands off from model-generated contract/spec to a deterministic compiler. "The mode is the factory in a sense." — Ricard Hansen: Arguing that defensibility lies in workflow and execution, not in the rented foundation model.

Implications: The episode suggests AI app builders may need to move from prompt-to-code toward data-to-spec-to-compiler systems, especially for enterprise internal tools. Products that rely on raw model output alone may be fragile, expensive, and easy to displace.