↓Skip to main content
  1. Projects/

Technical Publications Automation

·2 mins
Production AI Aerospace LLM Fine-Tuning RAG Enterprise
Table of Contents

Aircraft Maintenance Manual Automation #

Built a generative AI system to automate updates to Aircraft Maintenance Manuals (AMMs), converting complex technical documentation from unstructured PDFs into structured Markdown for LLM training and inference. This system addressed the challenge of processing 2,000+ Service Bulletins across 10,000+ page manuals.

Technical Challenge #

Standard PDF parsers failed on complex aerospace technical documents containing:

  • Multi-column layouts with jumbled text extraction
  • Tables spanning multiple pages
  • Technical diagrams and figures
  • Domain-specific terminology and formatting

Solution Architecture #

Custom PDF Parser #

  • Built from scratch using regex patterns for section extraction
  • Converted PDFs to Markdown preserving technical structure
  • Handled tables, diagrams, and technical content

Domain-Adapted LLM Training #

  • Full fine-tuning of GPT-2 and Mistral-7B models
  • LoRA adapters for aerospace vocabulary adaptation
  • Multi-stage instruction fine-tuning using Alpaca format
  • Processed on 4 A100 GPUs using PEFT and RLHF techniques

RAG System Implementation #

  • ChromaDB vector store for semantic search
  • LangChain agent for context retrieval
  • End-to-end training and validation pipeline

Impact & Results #

Operational Metrics #

  • Effort Reduction: 80% reduction in manual effort (90 days to 18 days)
  • Team Growth: Secured $200k funding, scaling team from 1 to 6 engineers
  • Process Automation: Complete automated document processing workflow

Recognition #

  • 1st place in Boeing LLM documentation hackathon
  • 3rd place globally in all Boeing hackathon categories
  • Validated solution across multiple aircraft manuals

Technology Stack #

Model Training

  • GPT-2 fine-tuning
  • Mistral-7B with LoRA/PEFT
  • RLHF for alignment
  • 4x A100 GPU training

Infrastructure

  • ChromaDB vector database
  • LangChain for orchestration
  • Custom PDF parsing pipeline
  • Azure ML infrastructure

Development

  • Python for ML pipelines
  • Hugging Face Transformers
  • PyTorch for deep learning

Professional Learnings #

Working in aerospace where errors have catastrophic consequences taught me the importance of rigorous validation, systematic testing, and building reliable systems. The experience of seeing a project from initial prototype to team expansion provided valuable lessons in scaling ML systems in enterprise environments.