The need
Find relevant code rules and tables from a plain-language job-site question.
Kory Kekewich← All projects
Electrical Assistant
A job-site question. A path back to the source.
A private electrical-reference app built for my brother, an Ontario apprentice, and used on real jobs.
At a glance
What I built
Source-linked AI chat, SQL lookup tools, twelve calculators, and equipment attachments.
My role
Product direction and AI-assisted development across data processing, backend, and phone interface.
Built with
- Next.js + React
- JavaScript phone interface and streaming API routes.
- Python + PyMuPDF
- PDF extraction and table-processing pipeline.
- PostgreSQL
- Supabase storage, full-text search, and typed lookup functions.
- Anthropic tools
- Bounded tool use, prompt caching, vision, and file attachments.
- Streaming
- NDJSON responses, saved conversations, and rule detail sheets.
- Verification tools
- SQL assertions and source-comparison tooling.
Product preview
Follow the sources
Rule and table chips link to supporting references, with a visible trace of the tools used to answer the question.
App screenshots shown in phone frames. A private reference tool, not a substitute for checking the governing code or qualified professional judgment.
Engineering decisions
Turning PDF tables into usable dataI combined Python extraction with batched vision transcription, repair passes, and side-by-side verification tools.
- Challenge
- Plain text extraction breaks the rows and columns that give a code table meaning.
- Outcome
- Reference content becomes searchable records and structured tables. Extraction checks do not certify every value.
Giving calculations a defined pathI built typed PostgreSQL tables and lookup functions that return supported values with source citations, plus known-value SQL assertions.
- Challenge
- A language model can misread tables or produce plausible but incorrect numbers.
- Outcome
- Dedicated tools handle supported lookups. The chat is not guaranteed to use them for every number, so answers still require verification.
Searching the way a person asksI combined exact-reference matching with PostgreSQL full-text search inside a capped tool-use loop.
- Challenge
- Questions arrive in everyday language, while the source is organized by exact rules and tables.
- Outcome
- The assistant can search, retrieve a rule, and request a lookup in one conversation, with references available to inspect.
How AI is used
Tools and references, not just a chat prompt
Claude chooses search and lookup tools within a bounded loop. Retrieval uses exact matches and PostgreSQL full-text search, not embeddings. References and calculator functions support checking; they do not guarantee correct advice. This private reference aid does not replace the official code or qualified judgment. Source-book text is not republished on this portfolio.
My contribution
Product direction and AI-assisted development across the ingestion pipeline, reference data, lookup tools, model integration, phone interface, and delivery for a specific apprentice's use case.
Where it stands
Built for one apprentice. Used on real jobs.
A private reference tool for my brother. Answers require checking against the official code.
Next step
Recheck database connectivity and run end-to-end reference and calculator checks before sharing a demonstration.