Hey y'all! We’re Oscar and Akshay from Saphira. We collaborate with robotics and aerospace companies to certify their products, which requires us to be deeply familiar with relevant standards and frameworks to produce requirements and test lists.
To automate our work, we built an AI agent that we’re making publicly available. It specifically searches for appropriately structured technical details in large, hierarchical documents. Here's a demo!
1. Understand Your Context
If you’ve ever had the “joy” of digging through ISO, UL standards, or OSHA/FAA CFRs while preparing for hardware product certification, you know how chaotic it can get. Many documents reference dozens of sub-frameworks, creating a complex web of standards.
We encountered this firsthand—and set out to fix it.
2. Choose the Right Approach
We learned by writing safety cases and building compliance plans that our AI agent must be capable of:
- Identifying the exact section of a standard relevant to a specific engineering need
- Parsing details like insulation width of a wire to prevent heat and current transfer (touch safety!)
- Traversing a knowledge graph of linked standards and substandards
3. Key Design or Execution Principles
Targeted Query Resolution
Our agent is optimized to surface technical content—not just definitions or reference chains.
Standards-Aware Search
By linking standards semantically, we enable queries like: "Find the safety criteria for DC brushless motor mounting."
Custom Contextualization
You can upload arbitrary documents to our Context Store, making the AI search contextually aware of your project’s needs.
4. Mistakes to Watch Out For
- Getting lost in circular references between standards
- Misidentifying purely definitional sections as technically informative
- Failing to capture substandard dependencies, which breaks traceability
5. Tools and Resources
Saphira helps teams:
- Log in and associate standards with their project
- Issue complex structured queries to find exact clauses
- Upload project-specific documents for tailored AI assistance
We’re actively improving all of this—please try it out and share feedback!