Air-Gapped LLM Migration Pipeline
A federal financial-systems modernization program had to reverse-engineer a large legacy Oracle EBS estate into modern application designs — work that consumed weeks of senior engineering time per application.
Air-gapped environment with cloud AI prohibited outright, scarce labelled training data, and a laboratory AI-governance process that had to approve any model before it could be used.
A six-stage offline agentic pipeline — parse, schema mapping, overview, parallel page and inventory generation, assembly — running locally adapted open-weight models behind an automated evaluation gate.
Documentation effort cut from weeks to days, 5,000+ lines of implementation-ready design documentation delivered, and no checkpoint scoring under the 90% quality bar ever reached the pipeline. When governance ruled out the original model family, I re-architected onto an approved one without losing pipeline capability.
Technical deep dive
XMLTABLE parsing of OAF definitions into a structured intermediate representation; rank-16 LoRA adapters on the attention projections (~15M trainable parameters) trained under both HuggingFace Transformers on MPS bf16 and Apple MLX; Jinja2 slot-based prompt assembly to force deterministic structured output. Training data was the binding constraint, not compute — a scarce set of curated gold examples was expanded 7× with synthetic instruction pairs, and every checkpoint scored against a weighted benchmark (structure, label accuracy, sample-data fidelity, inventory completeness) before promotion. Ran on a Mac Studio M3 Ultra with 256 GB unified memory, packaged for air-gap transfer with reproducible offline installs.