Selected Work - Automation

BUC Populi - a manual weekly grind turned into a hands-off pipeline

Professors were losing hours every week to grading and a high-stakes attendance audit done by hand. I replaced the manual grind with a weekly cron pipeline-fetch, AI-draft, audit, email-that runs on its own. AI does the first pass at scale; a human approves every grade. Automation deleted the toil; it never took the judgment.

AUTOMATION· weekly cron pipeline
live

cron

mon 07:00

fetch

Populi API

grade

Claude

audit

attendance

email

SMTP

generate_weekly_report.py
$ cron · mon 07:00
→ fetch_populi  ✓ 12 courses · 326 submissions
→ grade (claude)  ✓ 48 drafts · pending_review
→ audit (rules)  ⚠ 3 attendance flags raised
→ email (smtp)  ✓ professors notified
done in 4.2s
submissions326
graded · draft48
attendance flags3
45/minPopuli requests, limit 50

Role

Solo Engineer / Automation

Focus

Problem · Manual · Solution · Automation

Links

Internal tool

01The problem, and the manual way it was handled

Two recurring tasks were eating a university alive. Grading written work took faculty hours they would rather spend teaching, and every week someone had to reconcile attendance and submissions to prove attendance compliance, a manual audit where one missed week is a real institutional risk. Both were done by hand: a professor read essays one at a time while an admin cross-checked spreadsheets and chased missing records. It is the textbook shape of work to automate, with high volume, high repetition, and a high cost of mistakes.

02What I automated

  • Scheduled fetch - a weekly cron pulls courses, submissions, and attendance straight from the Populi API-no one logs in to export anything.

  • AI-drafted grades - Claude does a rubric-anchored first pass on every essay-per-criterion scores, feedback, flags-held as a draft at pending_review, never submitted on its own.

  • Automated compliance audit - the pipeline reconciles attendance and submissions per week and raises exactly which students are at attendance-compliance risk - the audit nobody wanted to run by hand.

  • Hands-off delivery - an SMTP step emails each professor their report on a fixed cadence-the workflow runs whether or not anyone remembers it exists.

03Automation decisions

Automate the chore, not the call

Problem
The instinct is to automate grading, but faculty will never trust a tool that writes grades on its own, and the painful work was really the repetitive weekly audit.
Approach
I pointed the automation at the toil, namely the scheduled fetch, the first-pass draft, and the compliance reconciliation, and left the final academic call to the professor.
Result
Grading time drops and the weekly audit runs itself, without anyone giving up control of the decision.

The human gate is the feature

Problem
An AI grade flowing straight into a record is unacceptable academically and ethically, and it would kill adoption on the first day.
Approach
I made the approval gate the core of the workflow. Every AI output lands as a draft at pending_review, nothing is pushed back automatically, and the professor is the system of record.
Result
The AI does the volume and the human keeps authority, so faculty trust the automation enough to run it.

A pipeline, not a button

Problem
A tool you have to remember to open is just manual work with extra steps, and it gets skipped the week it matters most.
Approach
I wired the whole thing as a weekly cron and email rhythm so the report arrives without anyone triggering it, and made it idempotent enough to re-run safely.
Result
The compliance check happens every week by default instead of depending on a busy person remembering to do it.

Compliance derived, not eyeballed

Problem
Attendance compliance is easy to get wrong by hand, since hybrid weeks, timezone-edge submissions, and per-course quirks make spot-checking unreliable exactly where it is high-stakes.
Approach
The audit derives week type and presence straight from Populi instead of assuming a schedule, computes compliance, and surfaces only the exceptions a human needs to see.
Result
A defensible, repeatable weekly audit derived from the source of truth rather than eyeballed.

04How it was built

Deliberately small and sharp, with the LMS as the source of truth, not a second copy.

Python 3.12Populi API v2Anthropic ClaudeStreamlitSMTPcron

05Outcomes

45/min

Populi requests, under its limit of 50

Weekly

Compliance audit, now automatic

Human-approved

No grade ever auto-pushed