Back to Work

Vellumark: evidence-first AI for SR&ED claim preparation

A guided claim-preparation cycle for small businesses doing eligible R&D.

A pre-launch AI-assisted workflow that helps small business owners collect technical evidence, organize it against SR&ED requirements, and turn it into a reviewable claim draft.

Role
Product and evaluation design
Organisation
Maple Spark Labs
Timeframe
2025–present
Status
Internal development · Pre-launch
Kind
Product
Technology and standards
  • Generative AI
  • Large language models
  • LLM evaluation
Focus
  • SR&ED
  • Grounding and citations
  • Completeness evaluation
  • Evidence review
  • Human review
  • Pre-launch product
  1. Collect evidenceGather the plans, records and notes that describe the work.
  2. Organize the workArrange the evidence against SR&ED requirements.
  3. Draft with citationsDraft claim statements that point back to their source records.
  4. Check grounding and completenessTest whether statements are supported and whether the claim has gaps.
  5. Review uncertaintySurface what a person has to decide before anything enters a claim.
  6. Retain review historyKeep the review decisions with the claim.
The claim-preparation cycle being built and tested. This diagram shows the workflow and the questions its evaluation asks, not a production result or a validated benchmark.

Problem

Small businesses can do real experimental development and still struggle to prepare an SR&ED claim. The technical evidence is often spread across plans, commits, test results, notes, and people’s memories. Reconstructing that story at filing time is slow, and a weak evidence trail can make an otherwise eligible claim harder to support.

For an owner, the practical cost is more than paperwork. SR&ED funding can return money to the business for further product development, hiring, or the next experiment. The claim still has to be accurate and defensible.

Constraints

Claim material is private, and the final judgment cannot be delegated to an LLM. The workflow has to show which evidence supports a statement, identify missing information, and give a person clear review points before anything becomes part of a claim.

Approach

I am designing Vellumark as a facilitated claim-preparation cycle rather than an open-ended chatbot. It guides a business from evidence collection through organization, AI-assisted drafting, citation and completeness checks, and human review.

The system connects draft statements to source records, keeps model runs observable, and retains review decisions. Its evaluation separates three questions that are easy to confuse: Is the statement grounded in evidence? Is the claim complete against the task rubric? Is the result ready for a person to review?

Current state

Vellumark is in internal, pre-launch development. The guided workflow and evaluation suite are still being tested. No production benchmark, approval rate, or customer outcome is presented.

What I learned

A fluent draft is not a prepared claim. Evidence, completeness, and readiness for human review are separate product problems.

Related work