Mode A
Pre-hire, unauthenticated
Pick industry, sub-industry, department. We match you to a mature LoRA from that vertical. 15-minute session. No card, no sign-up. Email-gate to extend to 30 minutes.
The Interview Room
Try the agent first.
Hire them second.
Would you hire a human without interviewing them? Our Interview Room lets you spend time with any agent — even borrow a mature Au-grade agent from the hive — before committing them to your business. Throw your hardest scenarios, see them work, then decide.
Three interview modes
Pre-hire, authenticated, or your own.
Mode B
Pre-hire, authenticated
Authenticated tenants get longer sessions (60 minutes), transcripts saved to their account, and audit logging. Atom: 5 interviews per day. Proton and Fusion: unlimited.
Mode C
Ongoing testing
Interview your own deployed agents at any time — for prompt change testing, sanity checks, campaign validation. Sandboxed, doesn't pollute training data.
Borrow from the hive
Twelve months from now.
Today.
The Interview Room can serve you an aggregated representation of a vertical's current best agent — not a specific tenant's instance, but the collective capability the hive has produced. Four moving parts make it work:
Vertical LoRA registry
Each vertical has a current-best LoRA
Versioned, tagged with grade — e.g. `vertical/solar-residential:au-grade`. Stored in Cloudflare R2.
Shadow Hermes pool
Ten warm Hermes instances
Reserved on Hetzner for interview sessions. Each can load any vertical LoRA on demand.
Ephemeral context
Each session has its own memory
Scoped to that session. No long-term memory across interviews. Borrowed agents don't know each other's conversations.
Capability profile snapshot
On Hire, the capability transfers
Vertical LoRA + sub-agent topology + system prompt template become the starter pack for your new local instance.
Privacy & legal
Borrowed, not exposed.
Borrowed hive agents are aggregated representations of vertical capability — not specific tenant instances. No verbatim text from any tenant's conversations is retained in the trained LoRA.
Minimum 50 tenant contributors per vertical before its LoRA is eligible for borrowing. Pattern scans run pre-training to verify zero PII leakage. The borrowed agent's system prompt does not reference any tenant by name, sub-segment, or geography.