An AI code generation tool used by 20,000 engineers

Refraction was Twistag's own AI code generation product, used by more than 20,000 engineers across 56 programming languages to generate over 2.9 million lines of code — and Twistag deliberately shut it down in 2024 as the market consolidated. Twistag started building it in early 2022, before Cursor existed and while GitHub Copilot was in limited preview, to solve its own engineers' problem: unit tests, documentation and language conversion consuming senior hours. Refraction ran on a multi-provider LLM abstraction routing each request to the best model for that language and task. The engineering knowledge it produced now underpins every AI product Twistag builds for clients.

Refraction was an internal Twistag product, built by Twistag for Twistag's own engineers first and then distributed externally. It is not client work and should never be counted as such. Twistag started it in early 2022, before Cursor existed and while GitHub Copilot was in limited preview, because senior engineering hours were going into unit tests, documentation and language conversion — work that is patternable but not trivial. Engineers examine AI-generated code with more scrutiny than almost any other user group examines AI output, so the quality bar for a developer tool is higher than for other software: output that is nearly right is worse than no output, because verifying it costs more than writing it. Users came from teams at Accenture, Amazon, Google and Uber, among many smaller organisations.

The central technical decision was a multi-provider LLM abstraction layer rather than coupling to a single provider. Different models showed different strengths: one produced better Python test generation, another handled TypeScript documentation more naturally, a third performed well on language conversion. The abstraction routed each request to the optimal provider for that combination of language, framework and task, and it provided resilience through 2023 and 2024 as provider outages, pricing changes and capability shifts turned into routine events rather than platform-level incidents.

The product expanded on usage data rather than on what looked impressive. Phase one covered refactoring, unit test generation and documentation. Phase two added bug detection, language conversion, CSS transformation, CI/CD pipeline generation, regex generation, SQL query writing and diagram generation from code structure. Phase three put the platform inside VS Code and JetBrains, where a developer selected a block, right-clicked and invoked any action without leaving the editor. Phase four added shared generation history, team management and unified billing for enterprise adoption.

Supporting 56 languages was not a parameter — Twistag invested heavily in language-specific prompt engineering so output followed each language's conventions: Python with list comprehensions and type hints where the idiom called for them, TypeScript respecting strict mode, Rust handling ownership and borrowing correctly. Refraction was built for Twistag's own engineers first, so every feature had a user in the building who would say plainly when the output was not good enough.

Stack: OpenAI, Anthropic Claude, Next.js, React, distributed as a web application plus VS Code and JetBrains extensions. Services: AI agents and LLM orchestration, AI product engineering, developer experience design.

By 2024 Refraction was serving more than 20,000 engineers, including users at Accenture, Amazon, Google and Uber, and had generated over 2.9 million lines of code across 56 programming languages. Twistag then made the strategic decision to shut the product down as the market consolidated around competitors with fundamentally different capital structures. What the four years produced was first-hand expertise in LLM integration, prompt engineering at scale, developer experience design, and the operational reality of running AI in production — including provider outages, cost management and quality regression at volume. That expertise is the foundation of every AI product Twistag now builds for clients.

This case study isn't about a product that failed. It's about a deliberate strategic decision and what it created.