Work

Assista

An AI assistant and operations platform for Wood's product knowledge, customer support, escalations, observability, and knowledge-gap review.

AssistaAI product support / Internal operations

Temporary visual

AI support architecture for product knowledge and staff operations.

RAG answers
Agent traces
Escalations

Agent trace

observe

Intake

intent, room context, sentiment

Retriever

Postgres products + Pinecone chunks

Specialist

recommendation or troubleshooting route

Guardrail

scope, factual traceability, tone

7+

Agent services

RAG

Product answer flow

Observe / enforce

Guardrail modes

Admin

Knowledge-gap workflow

Stack

Next.jsNestJSSupabasePostgresPineconeOpenAI APIsTypeScriptSanity

Problem

Wood's product support needed more than a public chatbot. The team needed product answers grounded in verified documentation, staff escalation paths for low-confidence questions, traceable guardrail review, and admin workflows for unresolved knowledge gaps.

The platform also had to protect privileged services. Supabase service-role keys, model-provider credentials, Pinecone access, webhook secrets, and cron secrets all needed to stay behind the backend boundary instead of leaking into browser code.

Solution

Atero built a full-stack platform with a NestJS API, Supabase Postgres/Auth/Storage, Pinecone retrieval, OpenAI-compatible model calls, and a Next.js internal frontend.

The system routes requests through typed agent services for intake, recommendations, specifications, comparisons, troubleshooting, guardrails, escalation summaries, and knowledge-gap clustering. The chat path combines product records, document chunks, classification, prompt construction, and model completion while preserving the public response contract.

Admin and internal workflows were added for escalation ownership, transcript review, trace inspection, guardrail observe progress, per-agent cost breakdown, and knowledge-gap report review.

Outcome

The project delivered public and internal chat surfaces, retrieval-backed product support, role-gated admin tools, escalation queue ownership, session memory controls, agent trace observability, and a knowledge-gap workflow that turns unresolved questions into prioritized content work.

  • Public and internal chat sessions with stable response shape.
  • Routed specialist agents for recommendations, specifications, comparisons, and troubleshooting.
  • Guardrail outcomes persisted on assistant messages with observe and enforce modes.
  • Low-confidence paths that can capture unresolved queries and create structured staff handoffs.
  • Authenticated session memory with bounded summaries, disclosure, update limits, and deletion support.
  • Scheduled knowledge-gap processing with admin review and audit-backed action history.

Implementation detail

How the system was designed and shipped.

Assista was built as a full-stack AI support and operations platform for Wood's product knowledge, customer support, and internal review workflows.

Backend boundary for privileged work

A NestJS API owns model calls, Supabase service-role access, Pinecone retrieval, webhook secrets, cron protection, and role-gated staff operations. Browser clients only receive public response data and authenticated API results.

Retrieval-backed answers

Product data in Postgres is combined with Pinecone document chunks, classification, prompt construction, and OpenAI-compatible generation so product answers can cite the system's verified knowledge instead of relying on unsupported assumptions.

Agent routing and observability

Typed NestJS agent services handle intake, recommendations, specifications, comparisons, troubleshooting, guardrails, escalations, and knowledge-gap clustering. Agent trace rows capture metadata, latency, token usage, success state, and input/output hashes without storing raw prompts.

Delivery tracks

01

Public and internal chat sessions with the existing response contract preserved.

02

Recommendation, specification, comparison, and troubleshooting agents with typed contracts.

03

Guardrail observation and enforcement modes for traceability, tone, scope, competitor handling, and unsupported claims.

04

Low-confidence retry paths, unresolved-query capture, staff escalation summaries, webhook notifications, ownership, status changes, and transcript review.

05

Authenticated session memory with bounded summaries, use disclosure, update limits, and deletion support.

06

Knowledge-gap processing that clusters unresolved queries, produces prioritized reports, supports admin review, and records audit-backed actions.

Product surfaces

Public chat and embeddable support widget
Internal staff chat
Escalation queue and transcript review
Admin observability dashboard
Knowledge-gap report review
Swagger API documentation

Screenshots

Placeholder visuals until product screenshots are approved.

These temporary images mark the intended screenshot slots and should be replaced with sanitized product captures.

Customer-facing product support placeholder

Public assistant

Customer-facing product support

Escalation queue and staff handoff placeholder

Internal operations

Escalation queue and staff handoff

Guardrails, traces, and knowledge gaps placeholder

Admin insight

Guardrails, traces, and knowledge gaps