This tool significantly reduced manual coding time, enabling developers to focus on higher-value tasks.
Meet our client
Client:
Industry:
Market:
Technology:
Client’s Challenge
AdaCore, a leader in Ada programming tools for mission-critical systems, sought to boost developer productivity by automating repetitive coding tasks. Existing tools lacked intelligent, context-aware code completion, slowing development and causing errors.
Our Solution
We developed a proof-of-concept code completion tool using StarCoder and CodeGen, fine-tuned for Ada code generation. The models were trained on 30,528 Ada code files from The Stack, representing 2.4% of the original corpus, and optimized for fill-in-the-middle (FIM) tasks. To boost performance, we applied memory-efficient techniques like LoRA and QLoRA.
Client’s Benefits
The fine-tuned models outperformed baseline versions in Ada code synthesis, even surpassing larger models like GitHub Copilot in accuracy for Ada code. This tool significantly reduced manual coding time, enabling developers to focus on higher-value tasks.
