Find matching parts without digging through catalogs
A distributor wins accounts by answering one question fast: here is the part number I buy from your competitor, what is your equivalent? This client's lookup tool failed three or four times out of ten. We built a Neo4j product knowledge graph holding more than 155,000 products across four catalogs, a search-score-approve workflow where consultants sign off on every match and each decision becomes permanent graph data, hybrid text, vector and image search for parts named for machines, an English-to-graph query layer, and a consultant workspace over Salesforce.
The owner & the bottleneck
Where it started.
The cross-reference question ran through an internal lookup tool that failed three or four times out of ten. What followed was the real workflow: catalog PDFs, supplier websites, a colleague who might remember, and an educated guess a customer's production line would depend on. Nearly 78,000 verified equivalences sat in a workbook that had never been queryable. Part names like HHCS 3/8-16 x 1-1/2 GR8 ZP break keyword search because every supplier abbreviates differently. And the daily pipeline around the catalog lived in Salesforce list views, record pages and the tracking spreadsheets the team had built because nobody could see it in one place.
What we built
The system we put in.
A Neo4j knowledge graph where equivalence is a relationship, not a column: the distributor's catalog, a sister division's and two national competitors' products, connected to suppliers, categories and each other, loaded through crawlers, per-source normalizers and a spec-parsing pass that promotes material, thread size and diameter to queryable properties. A match-and-approve workflow that searches, scores and lays evidence side by side; a human approves or rejects, and either decision is written into the graph with confidence, approver and timestamp so it never has to be made twice. Hybrid search blending full-text, text embeddings, product-image embeddings and spec-aware matching. A natural-language layer where Claude translates a question into a graph query, corrects its own errors and answers from real records. A consultant workspace that groups every open job by milestone with aging indicators and writes changes back to Salesforce. Started with fasteners, then expanded to nine categories.
The numbers
What changed, measured.
Resources
Keep learning
Claude Code Plugin Evals: The Traps the Docs Don't Mention
We ran Claude Code's new plugin eval harness on a real plugin: 8 cases, 46 agent runs, about $12. Here are the three ways the results JSON misleads you, what the eval sandbox can and cannot reach, and the five-line hook that fixes the plugin-root variable Bash never sees.
The Benchmark Pointed at Itself
The first paired benchmark of our procedural-memory registry measured a deterministic share under 0.5% on one workload and about 55% on another, caught its own row-append step failing silently, and ran under two conditions that rule out any arm-versus-arm verdict. Here is what held up, what did not, and the six variables the next cohort holds fixed.
Your Subagents Aren't Using Your Skills
We counted how often our AI subagents invoked skills from a several-hundred-skill library: zero of fifteen dispatches. Here is why cold-started agents never browse the library, the three injection levers that work, the spike methodology that proved the hook surface, and the log-first measurement instrument now watching every dispatch.
AI Agent Memory: Why Our Systems Stopped Re-Solving the Same Problems
Most AI automations re-derive the same work every run. Here is how procedural memory, storing the working code rather than a description of it, makes AI agent workflows faster, cheaper, and more predictable.
Rhize Tasks: A Local-First AI Task Planner for Jira and Calendar
Rhize Tasks turns Jira work into a realistic daily plan across Google Calendar and Apple Reminders, with local-first privacy and human approval controls.
We gave our AI agent 500 skills. It needed a graph, not a list.
Flat lists of AI agent skills fail silently as they grow. Every failure we hit was a relationship failure, so we built a generated graph, and the case for generating it came from our own code drifting within hours.
What could this look like
in your business?
Start with a conversation.
Tell us about your operation and what you want to improve. We'll talk through whether a similar approach could help.
No pitch, no obligation.