Individual complaints get lost in review feeds, while recurring issues are difficult to trace back to specific evidence.
02
My role
I owned source selection, n8n orchestration, the Node worker and PostgreSQL state, controlled HubSpot and Slack delivery, recovery, privacy maintenance and evaluation.
03
The approach
The official Steam API supplies attributed Coffee Talk reviews. Self-hosted n8n orchestrates a private Node worker, while PostgreSQL preserves stable identities, revisions, effect claims and delivery receipts.
Keep the real business identity, Toge Productions’ Coffee Talk, and clearly label this as independent work rather than a client engagement.
Use stable review identities and atomic effect claims so unchanged replay does not duplicate external work.
Hold mixed or uncertain inputs for review. Hold unknown delivery outcomes for operator reconciliation before retry.
Keep every public response under human control, with no publishing adapter.
Follow one review into support work.
Dated evidence walkthrough, November 7, 2026. These controls change the
explanation only. They do not run a workflow or send anything.
Coffee Talk, by Toge Productions
An attributed Steam review enters the owned workflow through the official anonymous API. The business identity stays intact.
Source identity: steam:914800:237850686
Negative · Product · Story
Conservative deterministic rules make the feedback legible. Mixed, conflicting or suspicious inputs go to human review rather than receiving an unsupported confident label.
Authored evaluation: 22/25 matches; three negative misses held for review.
Repeated reads do not mean repeated tasks
PostgreSQL records the source identity and content version. Atomic effect claims serialize delivery. A changed review creates a revision without overwriting existing external work.
Later replay: no_pending_effect
A real task. An unposted draft.
HubSpot task 417943964402 contains the source link and a generic response draft. A person must verify the wording and publish separately. No fix or refund is promised.
UNPOSTED PUBLIC RESPONSE DRAFT · Developer-test destination
Two reviews, two overlapping themes
One Slack digest summarizes 33 observations. Product and story each refer to the same two negative reviews. The November 2–9 observation window was triggered early on November 7 and marked incomplete.
One verified digest · Not proof of a future Monday schedule firing
n8n coordinates, the worker decides, and PostgreSQL remembers.
Public publication stays outside the automation.
What happens if delivery is uncertain?
Tested recovery path: pending, then claimed, then uncertain. A
timeout may hide an accepted delivery, so the operator checks for a
receipt or confirms absence before resolving it. The system does not
blindly resend.
The worker conservatively tags sentiment and topics, creates HubSpot developer-test support tasks for negative reviews with generic unposted public-response drafts, and sends a recurring-theme observation-window digest to Slack.
05
Results & impact
The controlled run observed 33 real reviews, created two native HubSpot tasks and delivered one verified Slack digest. Later replays produced no pending duplicate effect. These are delivery observations, with no measured revenue lift or public response-time improvement.
06
Lessons & takeaways
Preserve ownership and human review across the connected workflow. Measure each outcome separately and retain the limits of the controlled evidence.
07
Explore the evidence
Source and scope
I selected the official anonymous Steam reviews API for Coffee Talk by Toge Productions. Attribution stays attached to the real source. This is an independent portfolio demonstration, not client work.
Operational boundaries
n8n coordinates execution. The worker owns normalization, conservative classification and destination interfaces. PostgreSQL remembers identities, revisions, claims and receipts. Uncertain delivery requires reconciliation before retry.
The digest groups first observations into a Chicago-time week. Its two negative topics refer to the same two reviews, not four complaints.
Validation and scope
22 unit checks and ten real PostgreSQL integration checks passed. Seven successful server n8n executions verified intake, delivery and replay. The classifier matched 22 of 25 authored cases, with three negative misses held for review and zero invalid outputs.
Coffee Talk is a real game from Toge Productions, not a coffee retailer or client.
The digest was triggered early on November 7 for the November 2–9 observation window and explicitly marked incomplete. A future Monday schedule and uninterrupted live coverage were not demonstrated.
Conservative deterministic rules matched 22 of 25 authored cases. This small evaluation is not a broad accuracy claim.
Intake and digest delivery are paused after validation. Privacy maintenance remains active. Every public response is human-published separately.
Next steps
Broaden the labeled classification evaluation and investigate the three negative misses.
Define a production dispatch batch policy and measure a complete monitoring period with agreed source access.
Record the prepared 4–5 minute hiring-manager walkthrough with reviewed captions and transcript.
