Role-scoped AI assistant
Five specialists in one chat
Per-specialist skill gating means an assistant cannot edit data outside its role. Forty-seven tool-bound skills, zero cross-domain hallucinations.
Read the caseAll case studies
Case study · Research specialist
Meridian Operations · read-only audit researchReconciliation in 30 seconds, not 8 minutes
The Research Specialist answers exactly one class of question: what happened between what we submitted and what came back? It cannot do anything else. That constraint is the feature.
0 sec
Average lookup time vs. 8 minutes by hand
0
Mutation actions — read-only by design
0%
Question coverage on the variance rubric
Built with
The problem
Two browser tabs, an eight-minute scroll, and a typed comment from memory
The team did variance investigations as a daily ritual. Pull the submitted record. Pull what came back. Compare by eye. Type the difference into the audit ticket. Most of the eight minutes per investigation went into mechanical tab-switching and value-by-value comparison. The cognitive part — was this difference expected? — happened in the last ten seconds.
- The eight-minute lookupFinding the difference between a submitted record and an upstream response meant opening two browser tabs, scrolling to the relevant fields, and comparing by hand. If the record had been resubmitted, three tabs. The comparison lived in someone's head, not in a diff.
- The copy-paste chainAfter finding the difference, the analyst typed the finding into a comment on the audit ticket — from memory, occasionally with a field name wrong. The audit ticket said something different from what the record said.
- Re-reading the same record twiceDifferent team members ran the same lookup on the same submission because there was no shared research surface. Two people could spend 16 minutes reaching the same conclusion with no way for the second to know the first had already done it.
- Write access in a research contextBefore role scoping, a research session used the same agent instance as edit sessions. An analyst running a comparison could accidentally call a mutation tool with the wrong argument. The accidental write was rare but non-recoverable on submitted records.
Active reconciliations · 11:42 am
2 dupes in progressMO-48217in flight3:42 elapsed
MO-48214tab-switching1:18 elapsed
MO-48211typing into ticket5:03 elapsed
MO-48209in flight0:47 elapsed
MO-48205re-reading8:12 elapsed
MO-48204re-reading7:56 elapsed
Two analysts · same question · two answersNo shared surface
The pipeline
From inbox to verified record in one pass
Six steps. Five of them are deterministic. The model only owns the planning step.
01Pick
Research specialist selected
User opens the assistant and picks the read-only research persona from the specialist grid; grant required to see the option.
02Whitelist
Read-only tools only
Tool list filtered server-side to lookup, variance compute, and source-span fetch. No mutation tools are present in the API request — not gated, absent.
Claude
03Question
Natural-language ask
User asks: 'What changed between our submission and the upstream response on packet 48217?' No field names required; the specialist knows the schema.
04Plan
Diff plan composed
Specialist plans the lookup: pull the submission record, pull the upstream response, run a typed-field diff across the shared field set.
Claude
05Fetch
Records and source spans pulled
Tool calls execute in sequence: submission record → upstream response → diff computed. Source spans loaded from blob storage when needed.
Postgres + Azure Blob
06Respond
Field-by-field diff returned
Matching fields omitted; differing fields shown with submitted value, upstream value, and field identifier. Output is paste-ready into the audit ticket.
Claude
Clear packet reference
Diff returned in under thirty seconds
Specialist resolves the packet, retrieves both records, runs the diff, returns a clean field-by-field comparison. Output is paste-ready into the audit ticket.
Ambiguous reference or missing upstream
Decline rather than guess
If the upstream response hasn't been recorded yet, the specialist says so explicitly rather than returning a partial diff. If the packet reference is ambiguous (multiple resubmissions), the specialist lists the timestamps and asks the user to confirm. It does not guess.
Validation review
Coverage per lookup type
The specialist handles five types of variance question. Each is scored against a weekly sample for first-pass-correct rate.
Field-level confidence
Pass 2 — Claude self-review
HighMediumLow
Field diffSubmission vs. upstream response
98%High
Historical lookupReference number search
97%High
Variance attributionWhich pipeline step introduced the change
92%High
Source-span retrievalOriginal document segment
95%High
Audit-trail recapSequence of events on a packet
66%Low
Routed to human review. Re-imported records can carry timestamps from the original submission, creating a false ordering. Fix is a tool that strips import-time timestamps before passing the sequence to the model — not a prompt change.
4 of 5 fields cleared the 0.85 threshold
model: First-pass coverageThe stack
Boring tech, glued together well
Each vendor handles what it's best at. Aisyst owns the orchestration layer in between.
Claude
Plans diffs and generates structured field-by-field responses
PostgreSQL
Submission records, upstream response records, audit history
Drizzle ORM
Type-safe queries against the records and audit tables
Azure Blob
Source-span retrieval for original document segments
Third-party logos are trademarks of their respective owners and appear here only to indicate integration.
Outcomes
What changed when the research surface became shared
Two analysts asking the same question now get the same answer. Neither of them can accidentally write.
0 sec
Average lookup time
0
Mutation actions — read-only by design
0%
Coverage on variance-investigation rubric
0x
Faster than manual research
Watch rubric coverage on variance investigations
Currently 99%. If it falls, the specialist is encountering a new common variance pattern — a new upstream response format, a new field added to the submission schema, a new resubmission flow — and needs a new tool or an updated prompt. Coverage decay is earlier-warning than user complaints.