Two n8n workflows built for the Precision Data Solutions assessment, turning incoming emails and receipt photos into structured records and organized files.
What to inspect: recorded receipt vision-chain demo, test report, and importable workflows.
The implementation handles duplicate inputs, repeat writes, invalid extracted data, and branch-level failures through stable keys, idempotent upserts, validation, and explicit error reporting. JavaScript tests exercise the workflow Code nodes.
Scope: this is an assessment implementation. The recording verifies the receipt vision chain through a webhook harness; it does not establish a full production Telegram/Google deployment, extraction accuracy rate, or throughput benchmark.
- Workflow 1 — Multi-Label Gmail to Google Sheets & Drive. Polls a Gmail inbox on a schedule for incoming mail under multiple labels, logs sender details to a Google Sheet, and saves attachments to Drive with their original filenames, grouped in folders by label / date received / sender (33 nodes).
- Workflow 2 — Telegram Receipt Photo Processing. Accepts receipt photos via a Telegram bot, extracts merchant / amount / date (and more) with a vision model, logs them to a Google Sheet, and saves the original photo to Drive grouped by submission date (27 nodes).
Both JSONs import cleanly into n8n 2.37.10 and follow the same engineering standard: Config node for all IDs/thresholds, dedupe on stable keys, idempotent upserts, per-branch failure replies, and an Errors-tab logger — zero secrets in the workflow files.
A real execution inside n8n 2.37.10 (Docker), recorded end-to-end — webhook in, photo downloaded, vision extraction, validated record out in 11.2 s in one observed run. This is not a latency or throughput benchmark:
https://github.com/pihuts/pds-n8n-portfolio/raw/main/docs/images/demo_run.mp4
The demo harness (demo/demo_receipt_vision_chain.json) mirrors W2's hot path
with a webhook instead of the Telegram trigger and GLM-5.3-flash through an OpenAI-compatible endpoint instead of W2's Gemini endpoint — Photo to Base64 is the
verbatim W2 code and the parser applies W2's exact validation. What the run
produced from a real receipt photo:
| field | value |
|---|---|
| merchant | JOLLIBEE - ALABANG TOWN CENTER |
| amount | 468.16 PHP |
| receipt_date | 2026-09-07 |
| category | Food |
| items | Chickenjoy w/ Rice; Jolly Spaghetti; Pineapple Juice (Reg) |
| confidence | 0.95 |
| needs_review | false |
More evidence: executions tab (Succeeded in 10.99s), raw model reply + prompt, W1 canvas, W2 canvas, and the full write-up in docs/TEST_REPORT.md → "In-engine verification".
workflows/
workflow1_gmail_multi_label_to_sheets_drive.json W1: Gmail -> Sheets + Drive
workflow2_telegram_receipt_gemini.json W2: Telegram -> vision -> Sheets + Drive
demo/
demo_receipt_vision_chain.json in-engine demo of W2's vision chain (webhook-triggered)
docs/
SETUP.md step-by-step: credentials, IDs, sheet tabs, activation
SHEETS_SCHEMA.md exact tab + header layouts to create
ARCHITECTURE.md design decisions, data flow, idempotency, failure handling
TEST_REPORT.md what was tested, how, results, known limits + in-engine run
SUBMISSION_EMAIL.md ready-to-send reply for the HR thread
SUBMISSION_EMAIL.txt short plain-text version of the same reply
images/ screenshots + demo_run.mp4 (live execution recording)
config/
.env.example template for the single secret (vision API key, added as an
n8n Header-Auth credential - never stored in the workflows)
tests/
test_workflows.js structural validation of both workflow JSONs
test_code_nodes.js unit tests executing the real Code-node JavaScript
test_vision_spark.js live vision test: feeds a model reply through the real
Parse code and asserts the extracted receipt (5 checks)
fixtures/ sample receipts used in the tests
- Read
docs/SETUP.md(15-20 min, mostly clicking through the Google and Telegram consoles). - In n8n: Import from File, pick a workflow JSON, fill in its Config node (Sheet ID, Drive folder ID), connect the credentials, Activate.
- Send a test email / receipt photo and watch the first execution.
To reproduce the in-engine demo: import demo/demo_receipt_vision_chain.json,
click Execute workflow, then POST
{ "image_url": "<public URL of a receipt image>", "go_key": "<your key>" }
to /webhook-test/receipt-demo. This historical harness carries go_key in webhook and item data; execution retention can therefore persist it. Use a disposable key and a local test instance with execution retention disabled, then remove retained test executions. A reusable deployment should move authentication to an n8n credential. To run W2's chain against a different vision model, only the URL,
model name, and reply shape in the demo's vision nodes need changing; W2
itself ships pointed at Google Gemini with the key in a Header-Auth credential.
No secrets are stored in this package. The Gemini API key is added in the
n8n UI as a Header-Auth credential (header name x-goog-api-key) and
selected on the Gemini node - works on n8n Cloud and self-hosted, no env
vars required. The Telegram token lives in the n8n Telegram credential. The
historical demo workflow takes its key per request and builds the
Authorization header at runtime. The exported JSON contains no key, but the supplied
go_key enters execution item data and may be retained by n8n. Do not treat the demo
harness as a production credential-storage pattern.
Latest offline check: 29 September 2026 — structural checks pass; 45 code-node assertions pass. These checks do not execute the external services or measure extraction accuracy.
Run from the repo root (Node.js 18+, no extra packages needed):
node tests/test_workflows.js # structural checks
node tests/test_code_nodes.js # unit tests for the Code nodes
node tests/test_vision_spark.js <model-reply-file> # live vision chain
Provided as assessment evidence for Precision Data Solutions.

