PROJECT / 01 · AI WORKFLOWS

AI Support Ticket Automation

Classify fictional support requests once. Persist the outcome and its audit together, with accountable human-review decisions.

Locally verifiedLive on Oracle CloudReal AI classification with Ollama

INTERACTIVE DEMO

Try the support-ticket automation.

Submit a fictional ticket to see real AI classification, saved status and audit history. Runs on Oracle Cloud.

HISTORICAL TEST SNAPSHOTS

Testing & Reliability

Browse 13 historical snapshots. Each preserves the original execution date and recorded outcomes, including PASS and ERROR. These results describe past runs. Isolated, mocked, and infrastructure checks are labeled separately.

Local validation service must be available. Runs only isolated and mocked checks; no ticket or workflow actions.

Loading stored test evidence…

What I built

I connected n8n, JavaScript, MySQL and Ollama into a support-ticket pipeline with validated inputs, safe retries, transactional audits and authenticated human-review decisions. This demo submits synthetic tickets through the existing isolated automation.

Input → protected API → n8n → Ollama → MySQL result and audit.

THE BUSINESS PROBLEM

Support queues need consistent triage without duplicate processing, lost outcomes or unaudited manual decisions.

ARCHITECTURE / DATA FLOW

A clear path from input to outcome.

  1. 01

    Validate

    A protected backend bounds fictional input and assigns a synthetic ticket ID.

  2. 02

    Claim & classify

    n8n claims the ticket in MySQL before asking Ollama to classify a new request.

  3. 03

    Persist & review

    A transaction saves the result and audit; unusable AI output waits for authenticated human review.

ENGINEERING CHOICES

Where reliability needed deliberate design.

Retries must not become duplicate work

A payload fingerprint and conditional claim replay identical requests and reject changed content.

A result needs an audit

Ticket updates and audit insertion share a transaction; failure tests check rollback.

A demo must not reach unrelated data

A separate schema, restricted account, synthetic identifiers and private webhook isolate the demo.

TESTING / VERIFICATION

What the recorded tests established.

Previously recorded project results, not tests rerun by this website. Website CI checks the portfolio itself and does not execute these automations.

Selected actual test cases from each recorded suite. Outcomes below are previously reported PASS, not fresh automation runs or a complete case listing. Current source was inspected to identify the inputs and assertions; there is no public raw-results archive.

Backend checks

34 checks in the recorded suite · 2026-11-11

Actual local Node HTTP server; injected upstream fixtures. No n8n, MySQL or model execution.

Source: tests/demo-api.test.mjs in the automation repository. Public source excludes private runtime configuration and raw execution evidence.

Valid fictional ticket

Previously reported PASS
Test input
A unique requestId, fictional=true and “For a fictional order I was charged twice. Please investigate.”
Expected
Accept the request and return only safe result fields.
Recorded result
Response keys matched category, reason, status and ticketRef.
How checked
assert.deepEqual(Object.keys(out).sort(), [category, reason, status, ticketRef])
Why it matters
Prevents private customer or infrastructure fields leaking to the browser.

Reject personal information and invalid input

Previously reported PASS
Test input
Messages containing an email, URL or phone; short/1,501-character text; client ticketId; missing fictional acknowledgement.
Expected
Validation rejects each input.
Recorded result
validateBody returned false for each listed fixture.
How checked
assert.equal(validateBody(changedBody), false)
Why it matters
Keeps the demo fictional and bounds submitted data.

Block foreign requests

Previously reported PASS
Test input
Foreign Origin, missing session cookie or invalid CSRF token.
Expected
HTTP 403 before accepting a ticket.
Recorded result
Each blocked request returned 403.
How checked
assert.equal(response.status, 403)
Why it matters
Prevents other sites from submitting requests through a visitor’s session.

Duplicate and conflicting requests

Previously reported PASS
Test input
Repeat one requestId/message; then reuse the ID with changed text.
Expected
Same request succeeds; changed content returns 409.
Recorded result
Identical retry returned 200; changed message returned 409.
How checked
HTTP status assertions; upstream persistence is a Map fixture, not MySQL.
Why it matters
Avoids changing an existing ticket through a retry.

Concurrent duplicate requests

Previously reported PASS
Test input
Two simultaneous identical POST requests using Promise.all.
Expected
One upstream call; both requests succeed.
Recorded result
Call counter increased by one; both responses were 200.
How checked
assert.equal(calls - before, 1); pair.every(x => x.status === 200)
Why it matters
Avoids duplicate model work when a visitor double-clicks.

Timeout and rate limit

Previously reported PASS
Test input
A slow local HTTP upstream with a short deadline; separately, two requests against a one-request limit.
Expected
Timeout 504 promptly; excess request 429.
Recorded result
504 within one second; 429 with Retry-After: 60.
How checked
Elapsed-time/status assertions and Retry-After header assertion.
Why it matters
Bounds waiting time and protects limited compute capacity.

Demo integration checks

14 checks in the recorded suite · 2026-11-11

Actual local demo backend, complete DEMO n8n workflow and MySQL. Real Ollama on normal classification; fault cases explicitly substituted.

Source: tests/demo-integration.test.mjs in the automation repository. Public source excludes private runtime configuration and raw execution evidence.

Real classification and durable audit

Previously reported PASS
Test input
Fresh fictional duplicate-billing ticket through the protected demo API.
Expected
200, BILLING, AUTO_ROUTED and exactly two matching audit events.
Recorded result
Recorded 200/BILLING; saved message matched input; TICKET_CREATED and AI_CLASSIFICATION_ACCEPTED were present.
How checked
Real SELECT checked status, category, synthetic email, message bytes, auditCount=2 and ordered event types.
Why it matters
Proves classification is connected to durable storage rather than a frontend simulation.

Database isolation and private webhook

Previously reported PASS
Test input
Attempt original-schema table reads using demo account; POST directly without webhook authentication.
Expected
Original table access denied; unauthenticated webhook 403.
Recorded result
Both original tables were denied; direct webhook returned 403.
How checked
Permission probes plus assert.equal(response.status, 403).
Why it matters
Keeps a demo visitor away from unrelated records and private automation access.

Replay and conflict preserve stored data

Previously reported PASS
Test input
Repeat successful request; then same request ID with a different fictional message.
Expected
Replay returns original result without audits; changed content 409.
Recorded result
Replay deeply equaled first response; stored row unchanged after both requests.
How checked
assert.deepEqual(response, first); assert.deepEqual(row(ticketRef), saved).
Why it matters
Prevents retries or collisions from rewriting the original ticket.

Concurrent demo submission

Previously reported PASS
Test input
Submit the same new request twice simultaneously.
Expected
Both replies 200 with one persisted outcome.
Recorded result
Replies matched and the saved ticket still had only two audits.
How checked
Promise.all; response equality; independent MySQL auditCount=2 assertion.
Why it matters
Checks duplicate protection across the real backend/workflow/database boundary.

Uncertain classification goes to review

Previously reported PASS
Test input
Controlled classifier output: {"category":"UNKNOWN","reason":"Controlled synthetic model output"}.
Expected
202, PENDING_REVIEW and review-request audit.
Recorded result
Recorded 202/PENDING_REVIEW; TICKET_CREATED and HUMAN_REVIEW_REQUESTED saved.
How checked
HTTP assertions plus real MySQL status/event SELECT. Classifier node was temporarily substituted during the historical test.
Why it matters
An unusable classification must wait for human review instead of routing confidently.

Model failure is persisted and sanitized

Previously reported PASS
Test input
Controlled classifier error fixture, followed by repeating that failed request.
Expected
503 with safe error; FAILED and MODEL_UNAVAILABLE saved; replay does not retry AI.
Recorded result
Recorded 503/AUTOMATION_UNAVAILABLE; failed row/event sequence unchanged on replay.
How checked
Exact response comparison and before/after MySQL row comparison. This was not a real Ollama outage.
Why it matters
Reports dependency failure without leaking internal errors or silently repeating failed work.

Human-review logic checks

33 checks in the recorded suite · 2026-11-08

JavaScript executed with fabricated n8n inputs in a local VM. No workflow, SQL transaction or authentication endpoint execution.

Source: tests/review-decision.test.mjs in the automation repository. Public source excludes private runtime configuration and raw execution evidence.

Approval uses server identity

Previously reported PASS
Test input
Valid APPROVE request with BILLING and a reason.
Expected
Prepared status APPROVED with fixed server-side reviewer identity.
Recorded result
The isolated validator returned the expected status and server identity.
How checked
Assertions in “approval and fixed server identity”; this does not test Header Auth.
Why it matters
Stops the application from trusting a client-selected reviewer identity.

Reject invalid decisions

Previously reported PASS
Test input
UNKNOWN approval category, invalid/missing fields, malformed body fixtures.
Expected
400 for invalid decision input.
Recorded result
Recorded validation assertions returned 400 for the invalid fixtures.
How checked
assert.equal(validate(changedBody).httpStatus, 400)
Why it matters
Prevents incomplete decisions from reaching persistence logic.

Fail closed on inconsistent persistence output

Previously reported PASS
Test input
Fabricated database result with ticket update present but audit_rows=0.
Expected
Resolver returns 503 rather than claiming success.
Recorded result
Resolver returned 503.
How checked
assert.equal(run(resolveDecision, inconsistentRow, prepared).httpStatus, 503)
Why it matters
Avoids telling a reviewer that an unaudited decision succeeded; real rollback is tested separately.

Sanitize SQL error

Previously reported PASS
Test input
Fabricated error containing private driver detail.
Expected
Only safe REVIEW_UNAVAILABLE error and 503.
Recorded result
Returned exactly {httpStatus:503,errorCode:REVIEW_UNAVAILABLE}.
How checked
Deep equality of the entire returned object.
Why it matters
Keeps internal database details out of public responses.

Reviewer integration checks

20 checks in the recorded suite · 2026-11-08

Actual reviewer DEV webhook and MySQL with synthetic PENDING_REVIEW fixtures. Historical runs, not executed for this website change.

Source: tests/review-integration.test.mjs; docs/TESTING_GUIDE.md §3 in the automation repository. Public source excludes private runtime configuration and raw execution evidence.

Missing or wrong authentication

Previously reported PASS
Test input
APPROVE body with missing authentication, then intentionally invalid authentication.
Expected
403; ticket and terminal-audit count unchanged.
Recorded result
Recorded 403 for both; ticket remained PENDING_REVIEW, requires_review=1, UNKNOWN and zero terminal audits.
How checked
HTTP status assertions plus independent DB SELECT. Empty execution history alone is not proof of unchanged data.
Why it matters
Prevents unauthenticated ticket decisions.

Approval commits ticket and audit together

Previously reported PASS
Test input
Synthetic pending ticket, decisionId=decision-1, APPROVE, BILLING and a synthetic reason.
Expected
200; APPROVED, review flag cleared and one terminal audit.
Recorded result
Recorded 200/replayed=false; APPROVED,0,BILLING,1. Durable audit reference 11282.
How checked
DB SELECT and auditDetails checked fixed reviewer, previous/new status and matching reason.
Why it matters
Makes the accepted decision accountable and durable.

Identical decision is replayed

Previously reported PASS
Test input
Repeat exactly the committed approval.
Expected
200/replayed=true without a second terminal audit.
Recorded result
Recorded replay=true and exactly one terminal audit.
How checked
HTTP replay flag assertion and DB count.
Why it matters
Makes browser/network retries safe.

Two competing decisions

Previously reported PASS
Test input
Two simultaneous decisions for one pending ticket with different IDs/categories.
Expected
One 200 and one 409; exactly one terminal audit.
Recorded result
Recorded status set [200,409] with one terminal audit.
How checked
Promise.all; sorted status assertions; real DB count.
Why it matters
Ensures only one decision wins a race.

Audit insertion failure rolls back the ticket

Previously reported PASS
Test input
Synthetic pending fixture; historical test temporarily made the required audit event invalid with SQL NULL.
Expected
503; ticket remains pending with no terminal audit.
Recorded result
Recorded PENDING_REVIEW,1,UNKNOWN,0 after 503. Restored SQL then accepted the same decision with 200.
How checked
Independent SELECT after failure and recovery request. Final audit rows alone cannot reconstruct the transient rollback observation.
Why it matters
Prevents an approved ticket existing without its mandatory audit trail.

Reviewed tickets replay through intake

Previously reported PASS
Test input
Replay original synthetic ticket fields after APPROVED and REJECTED decisions.
Expected
200 with terminal status; no new classification or outcome write.
Recorded result
Recorded APPROVED and REJECTED replies; execution references 11297 and 11298.
How checked
Saved intake node data checked fields, no Classify Ticket/Save Ticket Outcome, and zero claim audit insertion.
Why it matters
Prevents a customer retry from reclassifying a completed human decision.

Source: Project README, root handoff and Milestone A testing guide. Local documentation reviewed 2026-11-12. Counts are separate suites, not a combined coverage score. Raw execution payloads and private receipts are excluded.

Read the evidence boundaries ↗

KNOWN LIMITS

What this does not claim.

  • The interactive demo runs on Oracle Cloud and accepts fictional tickets.
  • Actual service outages and connection loss during commit remain untested.
  • Reviewer/demo node histories are intentionally unsaved; durable audits and recorded output supply evidence.