Case Study: Agentic RAG

Transition Readiness Agent

A state-managed, multi-source RAG pipeline synthesizing signals from 8+ enterprise systems to prepare CSMs for seamless, auditable customer transitions.

System: multi-source synthesis

Input fragmentation

The handoff existed across systems, not in one usable view

When Professional Services finished onboarding, CSMs had to reconstruct objectives, commitments, sentiment, and risk before they could take responsible ownership of the account.

That reconstruction duplicated work, produced inconsistent transition messaging, and allowed hidden risks to surface only after the moment for proactive intervention had passed. The problem was not missing data. It was evidence without a governed sequence.

Fragmented evidence

Signals lived across Salesforce, Gong, Zendesk, Totango, Jira, and linked documents rather than in one dependable handoff.

Governed access

Customer data required role-aware access and PII controls; a useful brief could not expose raw source material indiscriminately.

Workflow latency

A briefing had to complete in under 15 seconds to remain viable in a live customer-success workflow.

Operational scale

The system had to serve 1,200+ accounts without trading away grounding quality.

Governed state machine

Every transition has a named state and an exit condition

The architecture moved away from one opaque synthesis prompt and toward explicit retrieval, extraction, synthesis, and validation states.

Ingest account context

Parallel retrieval gathers opportunity data, conversation signals, support history, and open delivery risks under the user's existing access.

Extract the contract

Given an Opportunity ID, a dedicated step locates the linked SOW and Order Form, parses contractual details, and returns a strict JSON payload.

Synthesize the brief

The orchestrator resolves the inputs into confirmed objectives, current health, known risks, and alignment gaps — a brief a CSM can inspect rather than a raw transcript dump.

Gate the outcome

Faithfulness is scored against a golden dataset before delivery or CRM write-back. Failed checks route to human review instead of passing uncertain output downstream.

Design decision

Reliability was worth 200 milliseconds

The initial monolithic prompt hit context limits, hallucinated edge cases, and returned inconsistent structures. Moving to strict state transitions and JSON contracts cost roughly 200ms per briefing; in exchange, grounding accuracy measured against the golden dataset improved 40%, and hallucinations fell to near-zero in the production workflow.

Inspectable brief

The output shows its shape before asking for trust

The CSM receives a structured account view with objectives, health, risks, gaps, and the control state that determined whether the brief could be delivered.

Supporting simulation / inspectable brief

Simulated operating model: transition readiness

Explore how a multi-agent orchestration pattern can coordinate customer context, readiness checks, and structured transition briefings across connected systems.

Systems modeled

Salesforce / Totango / Zendesk / Gong / Provisio / Google Drive / Slack

Simulation disclosure: This didactic exhibit uses illustrative data to demonstrate an operating pattern. It is not historical production telemetry.

Sourced outcome

Claims land beside their publication status

The evidence receipt distinguishes publishable proof from disclosed supporting context, keeping the outcome legible without presenting simulation as historical evidence.

Evidence dossier

Transition Readiness Agent

Approved visitor-safe summary derived from the evidence ledger. Sanitized depictions remain separated from production telemetry.

Ledger source: docs/proof/evidence-ledger.md

Publication classes: Public, Sanitized depiction

RolePublic: Defined the evaluation harness and retrieval strategy, integrated output into daily CS tools, and established prompt/quality monitoring frameworks.
ConstraintsPublic: Joined disconnected sources, respected governance and PII controls, operated under a sub-15-second briefing target, and served 1,200+ accounts.
JudgmentPublic: Replaced a monolithic prompt with a state-managed, multi-step agent chain and explicit validation gates.
Visitor-safe evidenceSanitized depiction: The existing state-managed orchestration diagram depicts ingestion, contract extraction, synthesis, and a quality gate.

14 Days

Earlier warning window for renewal-risk accounts

40%+

Adoption growth before the workflow became the standard handoff path

<2%

Hallucination rate across 1,200+ generated briefs

My role

Sole architect and builder

I defined the evaluation harness, designed the RAG retrieval strategy, and integrated the final output into the CS team's daily operational tools. I also established the prompt libraries and quality-monitoring framework behind that adoption growth.

Connected systems: Salesforce / Gong / Totango / Zendesk / Jira / Google Drive / Confluence
Control patterns: State-managed orchestration / SOQL retrieval / Strict JSON contracts / Role-aware access / Faithfulness scoring / Human review

Back to All Work

Source & Artifacts

Product Requirements Document