Repository navigation
How Ontora's (a prestigious YC company) Interview-to-Map Architecture Works — A Technical Reconstruction #4206
sanoojcools
started this conversation in
Show and tell
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
My Best Guess at Ontora's Interview-to-Knowledge-Graph Architecture
Hi all — I'm Sanuj (14+ years HR leadership, now building at the AI/org-design intersection). Relatively new to open-sourcing my work, so bear with me.
I came across Ontora (YC P26) and tried to reverse-engineer their architecture from their public demo. This is purely a learner's reconstruction — no insider access, and I may have misread things. That said, what a three-person team has built in four months is genuinely impressive. An end-to-end pipeline with orchestration, voice + text interviews, a fixed ontology, and a feedback loop — that's rare depth for an early-stage demo.
My Reconstruction
(Architecture diagram below — my own interpretation, not official)
The part that fascinates me most is the fixed ontology (Company → Dept → Product → Employee → Account → Process → System → Automation). If I'm reading the demo right, this replaces pure RAG with structured queries — e.g. "Which finance processes touch NetSuite with high automation potential?" Beautiful for consistency, but I wonder how it handles matrix teams or schema drift.
Three Genuine Questions
1. Fixed ontology vs. dynamic schema — Did you start rigid or let it emerge? What's the breakage point?
2. Feedback loop echo chambers — The "Show in sessions" toggle suggests past insights feed future prompts. How do you prevent compounding bias? (e.g. Agent assumes Jira → Employee confirms → confidence inflates). Contradiction detection, temporal versioning, or event-log validation?
3. Perception vs. precision — Metrics like "$27,310/yr effort" from transcripts feel powerful but fragile. How do you flag estimates vs. ground-truth before they hit executive dashboards?
I don't have good answers to any of these. Would love to learn from those who've built similar pipelines — especially with LangGraph or Neo4j.
If anyone from Ontora sees this: kudos on the ambition, and I'd genuinely love to know how off-base my reconstruction is.
Full notes: https://github.com/sanoojcools/-ontora-architecture-study
All reactions