Can AI speed up service without hiding the human transition?
Automate conversations without losing commercial context or hiding when AI should step aside.
CASE FILE. 04 / AUTOMATION / AI
Auto Whats
Connects WhatsApp by QR code, organizes the CRM, and uses business-aware AI to respond, qualify leads, schedule, follow up, and hand conversations to people.
- Next.js
- Fastify
- PostgreSQL
- Redis
- Baileys
BRIEF. 01 / PROBLEM AND BOUNDARY
The product in one read.
A deliberate view of the current state. Metrics and the hypothesis remain editorial until public evidence can replace them.
- CONNECTION
- QR + REALTIME EDITORIAL LABEL
- OPERATIONS
- CRM + AI EDITORIAL LABEL
- CAPABILITIES
- 6 frontend / backend / data / realtime / ai / integrations
- PUBLIC DOOR
- YES VALIDATED DEPLOYMENT
SYSTEM TRACE. 02 / PRIMARY FLOW
Boundaries before frameworks.
Web console → Fastify API → Redis/PostgreSQL queues and state → WhatsApp session through Baileys.
- 01CONSOLE→
Operator workspace
Next.js / inbox and CRM
- 02SERVICE→
Conversation API
Fastify / policies and commands
- 03STATE→
Queues + storage
Redis / PostgreSQL
- 04CHANNEL
WhatsApp + AI
Baileys / model / handoff
intent → policy → authoritative state → observable effectEXECUTION TRACE / ILLUSTRATIVE
One path. Every decision in view.
An editorial sequence for discussing contract, authority, and effect. It does not represent real production telemetry.
- CHANNEL
Normalizes, correlates, and removes known redelivery.
message.received - SERVICE
Resolves conversation owner, policy, and AI eligibility.
reply | handoff | hold - STATE
Persists context and queues only the effect that remains valid.
work.scheduled - CONSOLE
Updates inbox, CRM, and decision reason for the operator.
conversation.projected
DECISION LOG. 03 / TRADE-OFFS
Every choice costs something.
Treat AI as a per-conversation controllable mode and keep human handoff as a first-class product path.
- 01 / CONTROL
Per-conversation AI mode
- GAIN
- The operator can take over without competing with automated replies.
- COST
- Every queue must immediately honor the authority change.
- 02 / STATE
Redis for transit, Postgres for memory
- GAIN
- Low latency does not turn ephemeral state into permanent history.
- COST
- Retries and reconciliation cross two operational layers.
- 03 / AI
Model proposes; policy decides
- GAIN
- Qualification and replies accelerate inside auditable boundaries.
- COST
- Context, validation, and metrics cost more than a direct call.
RELIABILITY. 04 / FAIL WITH CONTEXT
Failure mode. Signal back.
These are not published production metrics. They are risks the design must make observable and candidate signals for the next round.
- 01
The WhatsApp session disconnects during a support sequence.
- 02
A message is redelivered and duplicates a CRM or follow-up effect.
- 03
Automation replies after the conversation has been handed to a person.
- 01
Recovery time and stability of connected sessions.
- 02
Handoff reason, human correction, and workflow resumption.
- 03
Duplicated, delayed, blocked, or unconfirmed messages.
INCIDENT DRILL. 05 / CONTAINMENT → RECOVERY → PROOF
When the hypothesis breaks.
Editorial resilience exercises, not real incidents or production runbooks. Each drill turns a failure mode into a response that can be discussed and tested.
DRILL-01The WhatsApp session disconnects during a support sequence.SIMULATION / NATIVE
Response to: The WhatsApp session disconnects during a support sequence.
- CONTAIN
- Mark the channel unavailable, limit the queue, and prevent replies that appear confirmed to the operator.
- RECOVER
- Reauthenticate the session, resume from the last acknowledgment, and reconcile messages by channel identifier.
- PROVE
- Confirm a stable connection, a backlog drained once, and a conversation projected in the acknowledged order.
DRILL-02A message is redelivered and duplicates a CRM or follow-up effect.SIMULATION / NATIVE
Response to: A message is redelivered and duplicates a CRM or follow-up effect.
- CONTAIN
- Deduplicate by external identifier and reuse each derived effect's idempotency key.
- RECOVER
- Reconcile receipt, message, and CRM action before releasing any still-unconfirmed retry.
- PROVE
- Demonstrate one logical message, one CRM effect, and one confirmation despite multiple deliveries.
DRILL-03Automation replies after the conversation has been handed to a person.SIMULATION / NATIVE
Response to: Automation replies after the conversation has been handed to a person.
- CONTAIN
- Increment the authority version and invalidate jobs whose expected owner is no longer AI.
- RECOVER
- Project the handoff in the console, keep the human as sole sender, and discard stale proposals.
- PROVE
- Run interleaving tests and confirm no stale output crosses the authority change.
DESIGN EXERCISE / REVIEW AGAINST REAL TELEMETRY AND RUNBOOKS BEFORE OPERATING
EVIDENCE. 06 / WHAT CAN BE DEFENDED
Facts in view. Gaps too.
Expand quality metrics, context recovery, and operational guarantees for connected sessions.
- Next.js + Fastify
- PostgreSQL + Redis
- Baileys + Firebase
- Human support handoff
INTERVIEW MODE. 07 / DEEPEN THE CONVERSATION
Three points to open on the board.
Questions that connect the case to system design, product, and operations without turning a contextual decision into a universal rule.
- 01
How to ensure a handoff invalidates already queued work.
↗ - 02
Which context to send the model and which state never to entrust to it.
↗ - 03
How to observe quality without retaining sensitive conversations indefinitely.
↗
CASE FILE / 04