Updated August 27, 2026 · 8 min read · Written by ProposalAI

AI/ML Proposal Example (PDF Q&A, Not “We Do AI”)

A legal-ops intern wants questions answered from a folder of PDFs. A useful proposal talks about retrieval, citations, and what the model should not do.

Example job situation

A small firm has ~200 internal PDFs. They want a chat box that answers from those files and shows which document a sentence came from. They said the answers must not invent policy. They asked if you fine-tune or retrieve. Budget is limited; they already have an OpenAI key.

Weak / generic proposal

I am an AI engineer. I can build ChatGPT, custom GPTs, agents, fine-tuned models, and computer vision. I will create a state-of-the-art AI system for your company.

This will 10x your productivity. Let’s start training your model this week.

Why this is weak

  • Fine-tuning is the expensive, unnecessary path for 200 PDFs.
  • “10x productivity” is an unsupported claim.
  • It never mentions citations, which was the actual requirement.

Improved proposal

Hi — for 200 PDFs I would retrieve, not fine-tune: chunk the documents, embed them, and answer only with context plus a citation (filename + page if we can parse it). If retrieval finds nothing close, the app should say it does not know rather than guess policy.

I built a similar internal Q&A over product manuals where the failure mode was confident wrong answers. The useful work was refusal when similarity was low, and showing the snippet the model used.

You already have an API key, so v1 can be a small Python or Node service plus a simple chat UI. Fine-tuning would not teach the model your PDFs and would cost more than this budget needs.

Are the PDFs text-based or scans (OCR)? And is this internal-only, or will clients type into it?

Why this is stronger

  • It answers retrieve vs fine-tune with a reason.
  • It treats “do not invent policy” as a product requirement (refusal).
  • Scope matches the budget and existing key.

Personalization points

  • Never promise legal accuracy. You are building retrieval, not a lawyer.
  • Ask about data residency if they are in a regulated industry.
  • Confirm they are allowed to send PDF text to their chosen API.

Questions you could ask

  • Text PDFs or scans?
  • Internal users only?
  • Any documents that must never leave your network?

Advertisement

Related reading