Voice of Customer Synthesizer

A1 · Mock Data Agent

• Purpose: Create realistic, multi-source VOC data for demos/tests.

• Inputs: Seed topics, product areas, segments, locales, devices.

• Process: Synthesizes reviews, support tickets, NPS verbatims with metadata (region, device, customer segment, product area, timestamps, thread IDs).

• Outputs: reviews.csv, tickets.csv, nps.csv (consistent schema).

• Checks: Class balance (neg/neu/pos), segment/locale coverage, time spread.

Note: This workflow requires this agent to simulate data, but would function with real user data from any product/company. In other words, this agent would be deprecated in a real-world environment

A2 · Ingestion & Clean Agent

• Purpose: Make data trustworthy and analysis-ready.

• Inputs: Raw CSVs.

• Process: Schema normalization, PII scrub (names/emails/phones), language detection, deduping & thread stitching, null/length guards.

• Outputs: voc_canonical.parquet (one tidy table).

• Checks: PII recall %, dedup rate, % rows passing quality gates.

A3 · Classifier Agent

• Purpose: Label each record with product area, device, region, customer segment.

• Inputs: Canonical VOC.

• Process: Lightweight text/rules/ML hybrid; confidence scoring; fallback heuristics.

• Outputs: Labeled VOC with product_area, device, region, segment.

• Checks: Spot-eval vs. small labeled set; coverage & low “unknown” rate.

A4 · Aspect & Sentiment Agent

• Purpose: Extract aspects (e.g., checkout speed, returns policy) and sentiment with severity.

• Inputs: Labeled VOC.

• Process: Aspect extraction, polarity (pos/neu/neg), severity (1–5), quote capture; aggregates by aspect/area/segment.

• Outputs: aspect_sentiment.parquet, top quotes per aspect.

• Checks: Aspect precision on eval set; quote de-duplication; severity sanity.

A5 · Topic Cluster Agent

• Purpose: Discover themes + map to a stable taxonomy; watch for drift.

• Inputs: Aspect snippets & embeddings.

• Process: Clustering (e.g., HDBSCAN/K-means), taxonomy mapping, drift detection vs previous run.

• Outputs: topics.parquet with topic, centroid terms, support, examples.

• Checks: Silhouette/DBI, min-support, taxonomy coverage, drift flags.

A6 · Scoring & Prioritization Agent

• Purpose: Turn signal into ranked opportunities.

• Inputs: Topic metrics (volume, impact proxy, severity), effort estimates.

• Process: Compute RICE, value/effort, and scenario weightings (source weights).

• Outputs: opportunity_scores.csv with topic, impact, volume, rice, effort_weeks.

• Checks: Outlier guards; monotonicity sanity; sensitivity to weight changes.

A7 · Feature Proposal Agent

• Purpose: Propose actionable solutions for top topics.

• Inputs: Top opportunities + context (area, segment, quotes).

• Process: Generate Feature title, Problem, Hypothesis, User stories & Acceptance Criteria, Success metrics, Risks, Experiment idea.

• Outputs: feature_briefs.csv + per-feature details rendered in the report.

• Checks: Duplicate detection, scope realism (effort bounds), metrics clarity.

A8 · Roadmap Agent

• Purpose: Convert picks → epics & timing under capacity.

• Inputs: Feature briefs + scores + current plan capacity per quarter.

• Process: Prioritization rule: Effort ascending → RICE rank ascending → tie-breakers VE (desc), Impact (desc), Volume (desc); assign to quarters respecting per-quarter capacity; ensure Quarter appears on feature pages.

• Outputs: roadmap.csv, Roadmap PNG (legend upper-right), ICS/CSV (optional).

• Checks: Capacity not exceeded; dependency ordering; label completeness.

A9 · QA & Eval Agent

• Purpose: Prevent regressions and ensure readability.

• Inputs: All intermediate outputs + a small labeled set.

• Process: Layout QA (no overlaps, strict margins), pagination tests, eval on classifiers & sentiment, drift checks; fixups (wrapping, truncation with ellipses).

• Outputs: QA log; green/red gates for publish.

• Checks: Page-level overlap=0, tables fit, eval metrics ≥ thresholds.

A10 · Publisher Agent

• Purpose: Ship tidy artifacts for stakeholders.

• Inputs: All curated outputs.

• Process: Build PDF (exec-ready), DOCX (Google Docs-ready), images (Opportunity Matrix filtered to proposed features, Roadmap prioritized), transparency CSV (prioritization table), optional Loom script.

• Outputs:

  • VOC_Report_V05_Run_n.pdf

  • VOC_Report_V05_Run_n.docx

  • Opportunity_Matrix_V05_Run_n.png

  • Roadmap_V05_Run_n.png

  • Prioritization_V05_Run_n.csv

• Checks: Correct naming/versioning; asset presence; links/legends; readability.