← Back to selected work

LLM engineering · Modernization

Processing codebases beyond a model's context window

A multi-stage approach to processing large codebases with chunking, intermediate aggregation, validation, retries, and batch-oriented workflows.

01 / The problem

Challenge

Large repositories cannot be treated as one prompt. Continuity and correctness have to survive across many bounded processing steps.

02 / System view

Architecture at a glance

A simplified, public-safe model. Select a layer to inspect its responsibility.

Selected layer

Repository

Source code and documentation exceed one model context window.

03 / Engineering decisions

Approach

  • Split source and documentation into logical, meaningful units.
  • Carry forward intermediate representations and summaries between stages.
  • Use repeatable validation and failure handling around batch processing.

04 / Tools & skills

What I used

Tools

Node.jsNestJSPythonOpenAI APIsBatch processing

Skills

Context-window-aware processingChunkingIntermediate aggregationRetries

05 / Evidence

Outcomes

  • Resume-reported approximately 40% less manual engineering effort.
  • Resume-reported approximately 30% fewer defects.

Resume-reported outcomes: approximately 40% less manual engineering effort and approximately 30% fewer defects. Confirm attribution before launch.

06 / Reflection

What I learned

01

Logical boundaries preserve meaning better than arbitrary truncation.

02

Intermediate state turns an oversized prompt into an observable pipeline.