Meridian Software
Meridian Assist: AI support desk inside the product
An in-app support assistant that answers from approved help docs, escalates cleanly to humans, and keeps coaches in control.
SaaS
Challenge
Problem space
B2B customers needed fast answers without waiting for business hours, without inventing policy.
Business impact
Long first-response queues and agent time spent on repetitive FAQs.
Solution direction
Ground answers in help content only, cite sources, and escalate when confidence is low.
Solution
We embedded a retrieval-grounded assistant in the product with citation links, escalation to ticket queues, and admin controls over which docs the model can use.
What we shipped
- In-app support chat
- Cited help articles
- Escalate to human
- Admin doc allow-list
- Evaluation logs
Architecture
Retrieve first, answer second. No citation, no reply—escalate instead.
In-app chat -> Assist API -> Doc retrieval -> LLM + guardrails -> Ticket escalation -> Admin allow-list
Product surface
In-app chat widget with cited answers and escalate CTA.
AI services
Chunk retrieval, refusal rules, and evaluation logging.
Ops
Admin doc management and ticket handoff payload.
Engineering challenges
Stopping invented policy
Friction
General models answered beyond published docs.
Solution
Hard retrieval gate and refusal copy when no chunk clears the score threshold.
Clean human handoff
Friction
Escalations lost chat context.
Solution
Ticket payload includes transcript, cited docs, and customer account id.
Outcomes
First response
- Before
- Queue wait for routine FAQs
- After
- Grounded assistant for after-hours questions
- Impact
- Agents spend live time on harder issues
Answer trust
- Before
- Open chatbot risk
- After
- Cited help-center answers only
- Impact
- Fewer off-policy replies
Technology
Product
- Next.js
- TypeScript
AI
- LLM API
- Vector retrieval
Backend
- Node.js
- PostgreSQL
Building something similar?
Tell us about your product. We will reply with next steps and an estimate path.