Taiwan B2B SaaS: AI Codebase Assistant Reduces Time-to-First-Commit by 55%

A Taipei-headquartered B2B SaaS company with 80 engineers and a 6-year-old codebase was experiencing a growing onboarding problem. New engineers joining the team required 6–8 weeks before making their first substantive code contribution — a timeline that had increased from 3 weeks three years earlier as the codebase grew more complex. The CTO calculated that the extended onboarding period was costing the company approximately NT$2.8M annually in lost productivity per engineer cohort, in addition to the morale cost of new hires feeling unproductive.

The core problem: the codebase had grown to 840,000 lines of code across 14 repositories, with inconsistent documentation and an architecture that had evolved through four rounds of significant refactoring. Tribal knowledge held by senior engineers was the primary navigation mechanism — but senior engineers were already at 85% capacity on feature delivery and could not sustainably serve as onboarding mentors.

The engagement started with a documentation audit: which repositories had readable README files, which had inline code comments, and which had architecture documentation? The audit revealed that 3 of 14 repositories had no README at all, and only 1 had current architecture documentation. However, all repositories had git history — thousands of commits with messages describing intent, Jira ticket references, and pull request descriptions explaining decisions.

The approach: build a retrieval-augmented generation system that indexed three corpora — git history, Jira tickets, and code comments — and made them queryable in natural language via a developer-facing chatbot integrated into the team's existing Slack workspace. Engineers could ask 'Why is the authentication module structured this way?', 'What does the data_pipeline_v3 module do?', or 'Where is the payment processing logic?' and receive answers synthesised from the indexed sources.

A secondary component automatically generated first-draft README files for the 13 repositories lacking them, which senior engineers reviewed and edited. This created a quality-controlled documentation baseline that the RAG system could also query.

The system was deployed as a Slack slash command (/codebase [query]) and as a VS Code extension sidebar. The VS Code integration allowed engineers to highlight code and ask 'explain this function' or 'find similar patterns in the codebase', with responses synthesised from the same indexed corpora.

The index was rebuilt nightly using a pipeline that re-indexed changed files and new git commits. Engineers could mark answers as 'helpful' or 'needs improvement', with feedback routed to the senior engineers responsible for the relevant repository — creating a continuous quality improvement loop.

The underlying model was a self-hosted Llama 3.1 70B instance (deployed on the company's existing AWS Taiwan infrastructure) for data sovereignty, with embedding generation via AWS Bedrock (Titan Embeddings) and vector storage in a self-hosted Chroma instance. No code or documentation left the company's AWS account.

Measured 90 days post-deployment against the previous 6-month cohort:

• Time-to-first-substantive-commit: 6.2 weeks → 2.8 weeks (55% reduction)
• New engineer questions routed to senior engineers: Estimated 65% reduction (based on Slack thread analysis comparing pre/post)
• Senior engineer time on mentoring questions: Self-reported reduction from 12 hours/week to 4 hours/week (67% reduction)
• README coverage: 3 of 14 repositories → 14 of 14 repositories (100%)
• Developer satisfaction with onboarding (post-90-day survey): 3.1/5 → 4.4/5
• Annualised productivity recovery: NT$3.2M (CTO estimate based on reduced onboarding lag per cohort)