The Workflow

Customers send requests for product labels through a shared inbox. Each request includes the exact text they want printed on the label, usually inside an Excel, PDF, or Word (.docx) attachment. Before production can begin, that information has to be turned into a correctly formatted label proof and sent back for customer approval.

Before the Automation

Before I automated the workflow, someone had to open the attachment, manually transfer the requested text into BarTender, format the label, capture an image of the finished proof, draft the response email, and keep track of which requests had already been handled.

I first documented the process as a training guide after new employees struggled with the same workflow. Once the process was clearly mapped out, I realized most of the repetitive work could be automated.

The Problem

The process was repetitive, interrupted other work, and depended on manually transferring customer-provided text without changing it. A typo, missing line, altered lot number, or formatting mistake could create an incorrect proof that still looked finished enough to send.

How the Automation Works

1. Watches the shared Outlook inbox through Microsoft Graph
2. Detects new label requests and reads Excel, PDF, or Word (.docx) attachments
3. Extracts the customer-provided label text
4. Formats the text for the standard 3×2 label while preserving the original content
5. Runs validation and safeguard checks before creating anything customer-facing
6. Opens BarTender and generates the formatted label proof
7. Captures the finished proof as an image
8. Creates a draft reply in the shared mailbox with the proof included
9. Tags the original request with a queue number and the SEND PROOF category, and keeps it pinned in Outlook
10. Leaves the final visual review and send decision with an employee

Safeguards

Every request follows the same validation path. If a check fails, the workflow routes it to a person instead of continuing automatically.

1. 1
   Only valid requests enter the workflow

   A request has to come from the approved customer, arrive after go-live and look like a label request. Supported Excel, PDF, and modern Word (.docx) files continue through validation.

   ↳ Routed to a person (Manual Review)
2. 2
   The text has to match

   Line breaks can change so the text fits the label. The words and numbers must still match the customer's file exactly.

   ↳ Routed to a person
3. 3
   Every attachment passes first

   All checks run before BarTender opens or a draft exists. If one attachment in an email fails, the whole email stops.

   ↳ Routed to a person
4. 4
   Failed attempts clean up partial output

   A failed attempt removes its own files and any partial draft, then retries up to three times.

   ↳ Routed to a person
5. Send
   A person sends it

   The automation leaves a prepared Reply All draft and the saved label files. An employee checks the proof and clicks Send. Printing waits for the customer's approval.

Running the whole time: the automation keeps its place in the inbox across restarts, and holds that place while any request is still retrying, so each email gets handled.

Edge cases

• Legacy .doc or .xls files- Manual Review
• Linked or cloud attachments- Manual Review
• Ship-To mixed in with Bill-To or notes- Reads only the Ship-To section
• Missing description or label size- Stops and routes to a person
• Nonstandard label sizes- Kept off the standard 3×2 path
• Several attachments in one email- One queue number, files numbered .1, .2
• A retry that finds an older draft- Deletes it so only one proof exists
• Microsoft rate limits or outages- Waits as instructed, then retries
• Expired mailbox sync- Re-checks from the last sync with a 10-minute overlap
• Requests after 3 PM on weekdays- Late notice based on when the email arrived
• A request that already finished- Skipped, so it's handled once

The Output

CHICKEN THIGHS, BONELESS

ITEM: 48271

LOT: 3289230

NET WT: 40 LB

KEEP FROZEN

STORE AT 0°F OR BELOW

To orders@sample-customer.com

Subject RE: Label request, PO 4471

Hi Dana,
Please approve the proof.
[proof image shown here]

Results

Built with Python, Microsoft Graph, SQLite, pdfplumber, openpyxl, and BarTender. 36 automated tests run on every push. Final sending remains human-controlled.

▶ Try the Interactive DemoOpen five sample requests and watch where each one goes. View the Code on GitHubPortfolio-safe reconstruction with fictional data, a runnable demo, sample files, and tests.

Open a request and watch where it goes. Recreated sample data.

orders@sample-customer.com9:14 AM Label request, PO 4471 PO4471.xlsx orders@sample-customer.com10:02 AM Label request, PO 4472 PO4472.pdf orders@sample-customer.com11:37 AM Label request, PO 4473 (4×6) PO4473.xlsx orders@sample-customer.com1:20 PM Label requests, PO 4474 and 4475 PO4474.xlsxPO4475.pdf orders@sample-customer.com4:20 PM Label request, PO 4476 PO4476.docx

No request open Waiting

1. 1
   Valid request
2. 2
   Text matches
3. 3
   Every attachment passes
4. 4
   Proof and draft built
5. Send
   A person sends it

Open a request to send it down the route.

The Workflow

Each weekly billing cycle starts with two QuickBooks exports that describe different parts of the same picture: invoice detail and current open balances. They have to agree before customer-facing billing reports can be created.

Before the Automation

A coworker told me he was overloaded with work, including this weekly billing process. I asked him to show me how he did it so I could understand the workflow and see what could be improved. Over a few Teams meetings, I watched the process, documented it as a training guide, and mapped out the repetitive steps.

The manual process was entirely click-driven in QuickBooks: open each unpaid invoice one at a time to mark it for printing, re-select every invoice in Print Forms, save the batch as a PDF, run the Open Balance report, then write the email by hand. It took about two hours a week. The only check was a final review before sending, so catching a wrong charge, missing tracking number, or omitted invoice depended entirely on someone noticing it.

As I worked through the automation, I realized the process worked best with a person still in it. The automation took over the repetitive work, while review, judgment calls, and the final send stayed with him. His weekly billing went from about two hours to 10–15 minutes.

The Problem

The workflow combines financial data with business rules. A report can look polished and still be wrong if one rate, balance, address, or invoice status is off, so the automation needed to reconcile the data before generating anything.

How the Automation Works

The desktop tool loads both exports, reconciles invoice detail against open balances, applies billing rules, surfaces exceptions, and generates the two customer PDFs only after the run passes validation.

Safeguards

Billing rule failures stop the run

Wrong packing items, rates, amounts, billing periods, or totals that disagree prevent the PDFs from being generated.

Unrecognized addresses require review

A location the system has not seen before stays visible for a person to verify instead of being guessed.

Missing tracking and unexpected tax are flagged

The run can continue when appropriate, but the issue remains flagged so the reviewer knows exactly what needs attention.

Previously billed weeks require confirmation

Every run is recorded by billing week. Generating PDFs for a week that was already billed requires a person to confirm it first.

Billing run review3 invoices checked

Invoice 10482Packing $12.50 flatPass

Invoice 10483No tracking numberWarn

Invoice 10484Packing is $15.00, expected $12.50Stop

PDFs locked until 1 hard stop is fixed

The Output

Invoice packet, first page Open-balance summary

Results

Built with Python and Tkinter. 15 automated regression tests run on every push. Final sending remains human-controlled.

▶ Try the Interactive DemoFind the invoice the automation would stop. View the Code on GitHubRunnable source with fictional QuickBooks exports, 15 regression tests, and sample PDFs.

This demo shows one of the billing rules the automation checks before report generation.

Three invoices are ready for validation. Packing & handling is a flat $12.50 per invoice. Which one should stop the run?

Each card shows a few details from one invoice.

Pick one.

The Problem

Job searching is mostly filtering. New listings appear constantly, the same role shows up on several boards, old posts stay up for weeks, and plenty of listings miss on location, pay, or job type. Reading all of it by hand takes time, and it takes even longer when you're searching for more than one person.

I wanted a system that could turn that noise into a repeatable pipeline: collect fresh listings, apply a candidate's criteria, remove duplicates and obvious mismatches, score what remains, explain why a role fits, and deliver a smaller shortlist worth actually reviewing.

How It Works

1. Pulls listings from selected job sources
2. Loads a candidate profile with its own resume, target roles, exclusions, and location rules
3. Filters by recency, role fit, location, and candidate-specific requirements
4. Removes duplicate listings before delivery
5. Scores promising matches and explains the strongest reasons they fit
6. Sends the shortlist through Telegram

Example Output

🟢 STRONG MATCH · 86%

Instructional Designer

Willowbrook Learning Studio

📍 Remote, US

💰 $72,000-$82,000/yr

Why it fits

• direct title match: instructional designer
• training content development
• e-learning and curriculum work
• remote
• full-time
• pay matches target range

Source: SampleJobs10:24 AM

Built for Different Candidates

Each candidate gets their own profile with target roles, exclusions, pay expectations, location rules, and how recent a listing has to be. Collection, filtering, deduplication, scoring, and delivery are shared, so adding another person means writing a new profile, not building another bot.

Personal project, actively developing the matching logic and delivery workflow.

▶ Try the Interactive DemoRun the search for three different candidates.

Pick a candidate, then run the search.

Each profile has its own target roles and rules. The same nine listings come in every time, and the results change with the profile.