02 / Forward Deployed AIEnterprise MVP / Telecom

Field OperationsAI Helpdesk

My roleForward Deployed AI Architect

ContextField technician operations

System classForward Deployed AI

Portfolio evidenceArchitecture · decisions · production

01 / Brief

The problem

Field technicians had to navigate fragmented knowledge, job systems and back-office support channels while working on-site. The objective was not another chatbot; it was a single operational entry point that could understand role, job context and escalation needs.

02 / Constraints

Why it was hard

01The primary experience lived inside a technician-facing Android application with a supervisor web surface.

02Authorization depended on SSO claims, role context and controlled supervisor impersonation for training.

03Knowledge arrived from databases, Confluence, SharePoint, PDFs, Word documents, incident records and images.

04The system needed grounded answers, citations and safe escalation rather than unbounded model behavior.

03 / Architecture

System shape

The architecture is expressed as operational layers: experience and identity at the edge, bounded orchestration in the middle, governed data and tools underneath, and evidence across the entire path.

Field Operations / logical architectureSanitized portfolio view

01ExperienceTechMobile + supervisor UI

02Manager agentIntent, role, job context

03Domain agentsJobInsights, TechAssist, Supervisor

04Knowledge layerS3, extraction, search / vectors

05OperationsRBAC, citations, HITL, telemetry

04 / Decisions

The engineering judgment

Route by intent and authorization together

The manager selects a domain path only after establishing what the user is allowed to access.

Separate domain agents, shared operational context

Job data, technical guidance and supervisor workflows have different tools and policies but share one request envelope.

Ingestion as a repeatable pipeline

Source-specific extraction feeds normalized chunks, metadata, permissions and index updates instead of ad-hoc uploads.

Escalation is a designed outcome

When confidence, permissions or source coverage are insufficient, the system moves the technician to the right human workflow.

05 / Production

Hardening the system

01Permission-aware retrieval and source metadata propagation.

02Nightly batch ingestion with failure visibility and reprocessing.

03Citation requirements for technical guidance.

04Supervisor and technician behavior separated through SSO claims.

05Operational telemetry across routing, retrieval, tool use and escalation.

Forward deployed engineeringCustomer workflow mappingLangGraphMulti-agent routingRAGSSORBACAWSOpenSearchpgvectorDocument ingestionHuman escalationObservability

06 / Outcome

A blueprint for reducing support friction by placing role-aware intelligence inside the technician’s existing workflow.

Next case study / 03