Like many massive enterprises, we should navigate the sweetness and chaos of legacy code. In our case, a long time of SQL procedures and enterprise logic that underpin a platform able to dealing with over 3 million concurrent customers and tons of of micro code deployments per week. It’s a posh machine. Contact one half, and also you danger breaking 10 others. That’s why modernizing the codebase is each a technical problem and a human one. It requires empathy, belief, and the flexibility to make knowledgeable guesses.
Contained in the Innovation Engine
At bet365, the platform innovation perform was established to impress risk. We’re a small, specialised crew charged with exploring rising and future applied sciences. Our goal is to determine the place they’ll have the best affect, and assist the broader group perceive tips on how to use them meaningfully.
We’re enablers and ambassadors for change. Our work spans every little thing from product growth and cybersecurity to the way forward for the workforce. Our guiding mannequin is McKinsey’s Three Horizons of Development reimagined for innovation. Horizon 1 focuses on what we are able to implement as we speak. Horizon 2 explores what’s coming subsequent. Horizon 3 dares us to think about the longer term nobody is speaking about but.
This framework helps us steadiness ambition with pragmatism. It creates area to experiment with out dropping sight of operational worth, and it ensures our builders, architects, and stakeholders are all a part of the identical dialog.
When GenAI Met Builders
When GPT-4 dropped in 2023, every little thing modified. Like most within the tech world, we had been fascinated. Generative AI provided a tantalizing imaginative and prescient of the longer term crammed with quicker insights, instantaneous summaries, and automatic refactoring. However the pleasure rapidly gave option to doubt. We handed very succesful builders a robust LLM and mentioned, “Go for it.” The outcomes had been combined at finest.
They inserted code into the immediate home windows, stripped out context to avoid wasting area, and hoped the AI would perceive. It didn’t. Builders had been confused, annoyed, and, understandably, skeptical. They noticed the AI as a shortcut, not a companion, and when the output didn’t match expectations, frustration adopted. Many requested the identical query: “Why am I asking a machine to write down code I may simply write myself?”
What we discovered was profound. The issue wasn’t the AI. It was the connection between the AI and the particular person utilizing it. We had assumed that talent in software program engineering would routinely translate to talent in immediate engineering. It didn’t. Did we miss one thing? The purpose we couldn’t overlook was through the train, our builders had been finishing the duties persistently round 80% of estimated time. There was undoubtedly one thing right here. We simply weren’t positive what it was. So, we went again to fundamentals.
Vibe Coding and the Limits of Belief
There’s a brand new time period in developer tradition: “vibe coding.” It’s the place you throw a bit of code at an LLM, get a response, tweak it, throw it again. Iterate quick. Ship quicker. It’s stylish. It’s seductive. However it isn’t danger free.
With out a clear understanding of intention or context, vibe coding can rapidly turn into a recreation of trial and error. And when your system is as complicated as ours – many databases processing 500,000 transactions a second – “trial and error” isn’t adequate. We wanted greater than vibes. We wanted imaginative and prescient.
Context Over Content material
The turning level got here once we realized the true job wasn’t educating AI tips on how to write higher code. It was educating people tips on how to talk with AI. We discovered a brand new mantra: intention + context + element. That’s what the AI wants. Not simply content material. Not simply “repair this perform.” However: “Right here’s what this code does, right here’s why it issues, and right here’s what I want it to turn into.” This perception is essential.
Our builders, particularly these tackling essentially the most complicated, interdependent issues, tailored rapidly. They had been used to pondering deeply, offering rationale, and navigating ambiguity. They acquired it. They fed the AI what it wanted. They flourished. The distinction was mindset. We got here to name this phenomenon “the unreliable narrator.” Not simply the AI, however the developer. As a result of typically, the issue wasn’t that the machine acquired it incorrect. It was at instances that we weren’t clear on what we had been asking.
RAG, GraphRAG, and the Energy of Grounded Context
To construct dependable, human-aligned AI help we would have liked a option to floor what the AI was seeing in actual fact. That’s the place we noticed the facility of Retrieval-Augmented Era (RAG). RAG permits an AI mannequin to retrieve related context from an exterior supply – like documentation, system metadata, or a information base – earlier than producing a response. It’s quicker to implement and extra versatile than fine-tuning, making it ideally suited for dynamic, domain-intensive environments like ours. Builders can replace the information base with out retraining the mannequin, protecting outputs present and grounded.
However RAG has its limits. When a query spans a number of programs or requires reasoning throughout disconnected items of knowledge, conventional RAG, which relies on textual content similarity, begins to falter. That’s why we turned to GraphRAG, a extra superior strategy that makes use of a information graph to reinforce LLM outputs.
A information graph doesn’t simply maintain information, it encodes relationships. It captures how elements work together, the place dependencies lie, and what may break for those who change one thing. GraphRAG makes use of this construction to enhance prompts at question time, giving the AI the relational context it must reply with precision. That is very true in environments the place accuracy is important, and hallucinations are unacceptable.
As a real-world train, we checked out our SQL server property. We wished to construct a system that we may use to realize invaluable perception on how the system works.
To construct it, we began by parsing all our database objects together with tables, views, procedures, features, and so forth. into summary syntax timber (ASTs). Utilizing Microsoft’s ScriptDOM, we extracted key information and used them to assemble the preliminary information graph. We overlaid this with pure language descriptions to additional clarify what every aspect did, and added runtime statistics like execution frequency, CPU time, and browse volumes.
The outcome was a wealthy, relational illustration of our SQL property, full with contextual insights about how objects are consumed and the way they work together. Then we surfaced this intelligence to builders by way of three core instruments:
- A chatbot that lets customers question the system in plain language
- A visualiser that renders a 3D map of dependencies and relationships
- A Cypher executor for superior graph querying and evaluation
What’s essential to notice is that many of the system’s worth lies within the graph, not the mannequin. The AI doesn’t have to know every little thing. It simply must know the place to look, and tips on how to ask the precise questions. That’s the facility of grounding.
For us, GraphRAG wasn’t only a nice-to-have, it turned important. It helped us transfer from generic code help to one thing way more invaluable: a system that understands what our code means, the way it behaves, and what it impacts.
We’re not simply writing code anymore. We’re curating it. We’re shaping the intentions behind it. Our builders now have tooling to realize additional perception to turn into code reviewers, system designers, and transformation brokers at an skilled degree throughout enormous division spanning architectures. All from a easy interface permitting pure language inquiries That’s the true shift. The longer term isn’t about AI doing our jobs. It’s about reimagining what the job is.
The success of our code modernization program has little to do with algorithms and every little thing to do with angle. We needed to unlearn outdated habits, rethink our relationship with code, and embrace a tradition of curiosity. We needed to cease asking AI for solutions and begin giving it the precise questions. The know-how was the simple half. The individuals half, now that was the true breakthrough.
