Sone-071 -

| Step | Description | |------|-------------| | 1. Intent Extraction | Use a lightweight BERT‑based classifier (trained on 150k historic queries) to label intent: date_range, numeric_range, status, tag, custom_field. | | 2. Entity Detection | Run spaCy NER + custom regexes for amounts, dates (relative like “last month”, “Q1 2025”), IDs. | | 3. Filter Generation | Map intent+entities to filter JSON structures. | | 4. Scoring | Score each candidate with a logistic regression that factors: confidence from intent, entity match count, historical acceptance rate (per tenant). | | 5. Result Count Estimation | Issue a lightweight COUNT(*) query using the generated filter on the search index (cached for 30 s). | | 6. Feedback Loop | Store SUGGESTION_APPLIED or SUGGESTION_REJECTED events. Retrain the ranking model nightly. |

Deployment

| # | As a … | I want … | So that … | |---|--------|----------|-----------| | US‑01 | Product marketer | A “New Notification” wizard that asks for channel, audience filter, schedule (date‑time or recurring), and message template | I can launch a campaign without developer help | | US‑02 | Product marketer | Time‑zone aware scheduling (auto‑detect from user profile) | Recipients receive the message at a sensible local hour | | US‑03 | Product marketer | Dynamic token insertion (first_name, plan_type, last_login) | Messages feel personal, driving higher CTR | | US‑04 | Product marketer | Delivery throttling (max 3 per user per 24 h, 10 per week) | We avoid spamming and stay under compliance limits | | US‑05 | Customer Success | Audit log showing who created/modified a schedule, timestamps, and version diff | We can investigate any delivery issues or compliance queries | | US‑06 | Data Engineer | Metrics endpoint (/notifications/metrics) exposing sent, delivered, opened, click‑through counts per schedule | We can monitor campaign health in our dashboard | | US‑07 | QA Engineer | Test harness to simulate time‑zone based deliveries & throttling | Automated regression coverage | SONE-071

| Sprint | Focus | Key Deliverables | |--------|-------|------------------| | Sprint 1 (2 wks) | Foundations | - Create notification-scheduler repo (service skeleton).
- Define DB schema & migrations.
- Implement basic CRUD API (no throttling). | | Sprint 2 (2 wks) | Audience & Tokens | - Integrate with user-service for audience query building.
- Hook into template-service for rendering. | | Sprint 3 (2 wks) | Delivery Worker & Throttling | - Background worker that reads ready schedules, resolves audience in batches, applies throttling (Redis).
- Provider adapters (SES, Firebase). | | Sprint 4 (2 wks) | UI & Wizard | - React pages & wizard modal.
- Form validation, preview rendering. | | Sprint 5 (1 wk) | Auditing & Metrics | - Audit logging, Prometheus metrics, Grafana dashboards. | | Sprint 6 (1 wk) | Security & Compliance | - Role‑based access, GDPR purge hook, encryption checks. | | Sprint 7 (1 wk) | Testing & QA | - Unit & integration tests, Cypress suite, load test script (10 k deliveries). | | Sprint 8 (1 wk) | Release Prep | - Canary rollout to staging, monitoring, documentation, training for marketing team. | | Sprint 9 (1 wk) | Production Rollout | - Production deploy, post‑deployment health checks, hand‑off. | | Step | Description | |------|-------------| | 1

Total: ~12 weeks (3 months). Adjust based on team capacity. | # | As a … | I

| # | As a … | I want … | So that … | |---|--------|----------|-----------| | US‑001 | End‑user (any role) | to type a free‑text query and instantly see a list of suggested filters (date range, status, numeric ranges, tags, custom fields) | I don’t have to manually add filters; I can get precise results faster. | | US‑002 | End‑user | to preview how many results each suggested filter will return before I apply it | I can decide whether a filter is worth using. | | US‑003 | End‑user | to apply a suggested filter with a single click and see the results instantly | My workflow stays fluid; I’m not interrupted by extra steps. | | US‑004 | End‑user | to undo any automatically applied filter (single “undo” button) | I can experiment without fear of losing data. | | US‑005 | Power‑user | to pin frequently used filter combinations as “Saved Smart‑Searches” | I can reuse complex queries with a single click. | | US‑006 | Data‑science / Product analyst | to see aggregate metrics on suggestion acceptance/rejection rates per user segment | We can measure the usefulness of the feature and feed it back into the ML model. | | US‑007 | Admin | to toggle the Smart‑Search feature on/off per tenant | We can roll it out gradually or disable it for customers that need a strict UI. | | US‑008 | End‑user | to get natural‑language help (e.g., “What does “last quarter” mean?” tooltip) | I understand the implicit logic behind the suggestions. |