Khaqan. Shaheen
Home / Writing

Writing

18 pieces so far: articles, tutorials and working notes. All self-published here, all from work I have done, none with numbers I cannot show you how I measured. Subscribe by RSS. Guest bylines and talks will be listed on the press kit page as they happen.

Articles

Longer pieces on ERP, AI in operations, infrastructure and running an IT function. Written from work I have actually done. All articles.

Why I extended one ERP instead of buying a second system for manufacturing

A second manufacturing system means a second set of records and an interface to maintain forever; extending one Odoo instance kept six sites on one set.

Replacing local passwords with single sign-on in a mid-sized group

Identity first, then the application's own access model, then monitoring. That order moved a six-site group onto Google Workspace SSO with two factor.

Five AI systems in production: what breaks when nobody is watching

Connections drop, inputs drift and confident wrong answers get through. Running five AI systems in daily production taught me the rules that keep them safe.

Bringing a new factory online: the order of operations

Cabling and circuits first, then security, ERP, identity, telephony and site devices. The sequence that puts a new factory on group systems from opening day.

A PostgreSQL upgrade nobody was allowed to notice

Rehearse on a copy of production, keep a rollback inside the window, move one site at a time. PostgreSQL 9.5 to 16 with no unplanned downtime at any site.

Tutorials

Practical how-to guides with commands and code. Each one is a pattern I use, simplified so you can apply it. All tutorials.

Reading supplier bills into an ERP with an LLM OCR pipeline: a design that survives production

Design a supplier bill OCR pipeline with a vision model, strict JSON, validation rules, a review queue, idempotency and honest monthly accuracy checks.

Odoo: how to stop two work orders being booked on the same machine at the same time

Add a server-side overlap constraint to Odoo 17 work orders, back it with a PostgreSQL exclusion constraint, and give planners a drill-through to the clashes.

How to write an llms.txt for a business website

An llms.txt is a short Markdown map of your site at /llms.txt with an H1, a one-line summary and linked sections. Here is the format and a full worked example.

How to check whether ChatGPT, Perplexity and Google AI Overviews can read your website

Read robots.txt for GPTBot, OAI-SearchBot and PerplexityBot, fetch your homepage with curl as each bot, check the raw HTML, or run one Python audit script.

Google Workspace single sign-on for Odoo, with two factor authentication and role based access

Configure Google sign-in for Odoo 17 with auth_oauth, enforce 2-Step Verification in Workspace, map groups to job roles and keep a break-glass admin.

A WhatsApp AI sales agent that hands leads to your ERP: the architecture

A WhatsApp webhook, a session store keyed by phone, a model with four tools, a catalogue it cannot contradict, a hand-off rule and a lead created in Odoo.

A PostgreSQL major-version upgrade with no unplanned downtime: the checklist

Upgrade a production PostgreSQL database to 16 using logical replication, a rehearsed cutover and a rollback window, then rebuild backups and monitoring.

Working notes

Short entries from an engineering notebook. One observation each, from systems I run. All working notes.

Why every AI agent I run sits behind a human queue

Because a language model is confidently wrong in a way a scanner never is, every agent I run creates records for a person to check and none of them post or pay.

Why I published my case studies without numbers

Because a number I cannot show how I measured is worse than no number; the scope is honest, and measured figures will be added once I have measured them.

What running my own venture taught me about corporate IT

Paying the bill yourself, watching what customers do on a booking page, a pipeline that stops silently and search traffic answering back changed how I run IT.

What OCR gets wrong on supplier bills

OCR on supplier bills fails on handwritten totals, stamps over numbers, multi-page bills, duplicates and currency; the fix is validation and a review queue.

What I check the week before a factory opens

Two links with a tested failover, the company and warehouse in the ERP with local tax and numbering, printers on the floor, users logged in once, and a restore.

The month-end test for ERP health

If six sites can close the month from the same ERP records without a spreadsheet, the ERP is healthy; a spreadsheet at month end means a record is missing.