Cult.fit Moves Development to Codex: Social Feed in 2 Months Instead of a Year

A familiar picture: one team sticks with one assistant, another with a different one, and each has its own repository rules and its own prompts. Knowledge doesn't transfer, and reviews look different everywhere. Cult.fit, one of India's largest fitness companies (708 centers in 77 cities, almost a million paying members), solved this problem radically: it chose Codex as a single platform.

A caveat up front: this is a case study on OpenAI's website, the figures come from the company itself, and nobody has verified them independently. Still, the set of practices it describes is quite concrete and useful.

One Tool Instead of a Zoo

Before standardization, developers used a mix of agents and assistants. As a result, every team had its own conventions: repository structure, instructions for agents, skills, processes. Now everything is shared: structure, skills, testing, code review and feedback loops.

Specialized agents were built on top of this. An SRE agent pulls data from observability systems and logs, investigates an incident and in many cases creates a pull request with a proposed fix on its own, which an engineer then reviews. There are security agents that look for vulnerabilities, and on-call agents for recurring production problems.

The company is also moving toward spec-driven development: a specification carries a task through the entire lifecycle up to a PR that a human reviews. Example: the social feed in the app. CTO Santosh Kumar estimates it would previously have taken more than a year; the team shipped it in two months. That is his estimate, not a measurement.

Development for Non-Engineers

The second step is more interesting. The company made its data friendly to Codex, and teams began answering routine questions themselves instead of queuing for central analysts. HR built itself a performance review workflow: scheduling, feedback collection, consolidation of results. Over recent months, non-engineers have built more than 20 internal apps.

The dashboards are a story of their own: when an internal service suddenly became much more expensive, one engineer reproduced the needed functionality on Codex over a weekend. It now runs as an in-house app, saving more than $100,000 a year.

Note how this was rolled out. Access was not simply handed out. There was a large hackathon, then a one-day session for senior executives where each built something for their own function. Office hours are held twice a week, and engineers were embedded in individual departments for technical and domain support.

AI in the Product and Support

Through the OpenAI API, the company develops AI Trainer, used by about 20% of gym visitors. The new version is voice-based: built on GPT Realtime, the trainer offers tips during exercises and helps keep a workout log. Given that more than 45% of members came to the gym for the first time, betting on guidance makes sense.

In support, agents handle tasks that previously required a human. Example: a member asks for a fee to be waived for a missed class. The agent checks signals such as the weather in the city and attendance at that specific class, and makes a decision. The agents were embedded in the existing Sprinklr platform and connected to the backend. According to the company, support costs fell by more than 50%.

What to Take Away

First: standardization matters more than picking the best tool. Shared conventions enable skill reuse and predictable review. Second: if your bill for internal SaaS is growing, check whether it's easier to rebuild it yourself, but factor in maintenance costs, which the article doesn't mention. Third: training and embedded engineers, it seems, mattered no less than the model itself.

The limitations are obvious. There is no data on the quality and security of apps built by non-engineers, on the robustness of the agent's fee decisions, or on how much it costs to maintain these homegrown systems. This is vendor marketing material, so the figures should be taken as a guide, not a guarantee.