Fine-Tuned LLM Drafting Assistant for a Device Manufacturer's Quality Documentation
Let's Connect
Overview
What we built
A medical-device manufacturer with 10 production facilities had engineers drafting SOPs, CAPA reports and change-control documents entirely by hand, and formats drifted between sites. We fine-tuned a self-hosted LLM on the manufacturer's own approved documents so it generates structured first drafts that engineers edit rather than write from a blank page.
In plain terms: every quality document started from nothing. Engineers across 10 production facilities drafted SOPs, CAPA reports and change-control documents entirely by hand, and because there was no shared starting point, formats drifted between facilities even when the documents covered the same kind of work. First drafts took most of a working day to produce, and internal audits repeatedly flagged inconsistent structure and missing template sections, which dragged review cycles out across multiple rounds of correction before a document was finally approved.
We fine-tuned a self-hosted open-weight LLM on the manufacturer's corpus of approved quality documents and paired it with retrieval over current QMS templates and procedures, so the assistant drafts from the manufacturer's own approved style rather than a generic template. The drafting assistant sits inside the existing document workflow, generating structured first drafts that engineers edit and route through the unchanged review and approval chain, and no document is ever released without human sign-off. First-draft time for CAPA reports fell from roughly 6 hours to under 2, and average review rounds per document dropped from 3.4 to 1.8.
The Problem
Slow hand-written quality documents
Quality documentation at the manufacturer started from a blank page every time. Engineers across 10 production facilities drafted SOPs, CAPA reports and change-control documents entirely by hand, each one built up from memory and habit rather than a shared, current template. First drafts took most of a working day to produce, time that came directly out of engineering hours that could otherwise have gone to the work the documentation was describing.
Without a shared starting point, formats drifted between facilities. Two sites could produce a CAPA report that covered the same kind of issue in structurally different ways, and internal audits repeatedly flagged inconsistent structure and missing template sections as a result. Each flagged finding meant another round of correction, so review cycles stretched across multiple rounds before a document was finally approved.
The manufacturer already had the answer to consistent formatting, its own corpus of approved quality documents, it just was not being used as a starting point for new ones. Every engineer effectively reinvented structure and phrasing that had already been solved and approved somewhere else in the organisation, and the resulting rework across 10 facilities added up to a steady drag on review cycles.
Entirely manual drafting
Engineers drafted SOPs, CAPA reports and change-control documents entirely by hand across all 10 production facilities, with no shared starting point for a new document.
Format drift between facilities
Formats drifted between facilities because there was no common template being followed, so similar documents ended up structured in different ways.
Slow first drafts
First drafts took most of a working day to produce, time that came directly out of engineering hours across every facility.
Audit findings and rework
Internal audits repeatedly flagged inconsistent structure and missing template sections, dragging review cycles out across multiple rounds of correction.
What it was costing them
Every hand-written first draft cost most of a working day, and every audit finding on structure or missing sections meant another round of correction before a document could be approved. Across 10 production facilities, that added up to engineering time spent reinventing formats the manufacturer's own approved documents had already solved, and review cycles that stretched longer than the underlying quality work required.
The Solution
Fine-tuned drafting assistant
We fine-tuned a self-hosted open-weight LLM on the manufacturer's corpus of approved quality documents, so the model learns this manufacturer's actual approved structure and phrasing rather than a generic document format. Keeping the model self-hosted and open-weight mattered given the sensitivity of quality documentation in a regulated environment.
We paired the fine-tuned model with retrieval over current QMS templates and procedures, so a draft reflects whichever template and procedure actually apply to that document, not just whatever the model learned in training. The drafting assistant sits inside the existing document workflow, generating structured first drafts that engineers edit rather than starting from a blank page.
The review and approval chain did not change. Engineers edit every generated draft and route it through the unchanged review and approval chain, and no document is ever released without human sign-off. That meant adopting the assistant did not require renegotiating how quality documentation gets approved, only changing how a first draft comes to exist.
Key decisions
Fine-tune on the manufacturer's own corpus
The model was fine-tuned on the manufacturer's corpus of approved quality documents, so its drafts already reflect this organisation's approved structure and phrasing.
Pair fine-tuning with template retrieval
Retrieval over current QMS templates and procedures runs alongside the fine-tuned model, so each draft reflects the specific template that document actually requires.
Self-host an open-weight model
The LLM runs self-hosted and open-weight rather than through an external API, keeping regulated quality documentation inside the manufacturer's own environment.
Sit inside the existing workflow
The drafting assistant sits inside the existing document workflow rather than a separate tool, so engineers encounter it at the point where drafting already happens.
Keep the approval chain unchanged
Engineers edit and route every draft through the unchanged review and approval chain, so no document is ever released without human sign-off.
Measurable Impact
What changed after launch
Drafting got faster first. First-draft time for CAPA reports fell from roughly 6 hours to under 2, freeing engineering time that used to go into building a document's structure from scratch before the actual content could even begin.
Quality and speed improved together rather than trading off. Template-conformance findings in internal audits fell 63% over two audit cycles, and average review rounds per document dropped from 3.4 to 1.8, since drafts arrived already following the manufacturer's approved structure. All 10 facilities were onboarded to the assistant within 4 months, so the improvement reached the whole organisation rather than a single site.
First-draft time
Roughly 6 hours for a CAPA report
Under 2 hours for a first draft
Audit findings
Frequent template-conformance findings
Down 63% over two audit cycles
Review rounds
3.4 review rounds per document
1.8 review rounds per document
Facility rollout
Manual drafting at all 10 facilities
All 10 facilities onboarded within 4 months
Headline results
First-draft time for CAPA reports cut from roughly 6 hours to under 2
Template-conformance findings in internal audits down 63% over two audit cycles
Average review rounds per document reduced from 3.4 to 1.8
All 10 facilities onboarded to the assistant within 4 months
Tech & Tools Used
What powered the build
Every tool below earned its place in this engagement. Here is the part each one played.
Python (FastAPI)
Serves the drafting API that receives a document request and returns a structured first draft into the existing document workflow.
PyTorch
Runs the fine-tuning process on the manufacturer's corpus of approved quality documents, producing the model weights the drafting assistant runs on.
Hugging Face Transformers
Provides the model architecture and tooling used to fine-tune and serve the open-weight LLM on the manufacturer's own approved documents.
LoRA (PEFT)
Fine-tunes the model efficiently on the manufacturer's document corpus without retraining the full model from scratch for each update.
vLLM
Serves the fine-tuned model's inference requests, generating structured first drafts quickly enough to fit inside engineers' existing drafting workflow.
pgvector
Indexes current QMS templates and procedures so the retrieval step can pull the specific template a document actually needs to follow.
PostgreSQL
Stores document drafts, template metadata and review history, tracking each document from generated first draft through to signed-off approval.
React
Builds the drafting interface where engineers review, edit and submit generated drafts into the unchanged review and approval chain.
Keycloak
Manages authentication and access control for the drafting assistant, keeping regulated quality documentation restricted to authorised engineers.
Docker
Packages the fine-tuned model and drafting services so the same environment runs consistently across all 10 production facilities.
Ready to Build your Life Sciences & Med Devices Business with LLM Integration & Fine-Tuning
Ask Byte
Ask Byte
Typically replies instantly
just Now
Hi! I'm OrganByte's assistant. How can I help you today?
AI-generated content may be incorrect
