Things I’ve built

The Northlane agent organisation

26 role-specialized AI agents that do the bid, estimating and compliance work of a small company.

Each agent is a written role with its own memory: one per client account, plus an estimator, a product catalogue, a company-records keeper, an approvals specialist for the aviation regulator and a chief-operating-officer agent that routes requests to the rest. I design the roles, the hand-offs and the rules each one follows, and Claude Code runs them. Smaller models sort and extract, the strongest model judges, and a person approves every price and every outbound document.

What came of itAbout 90% of my own operating workload now runs on the system (an operating-model figure, not an audited metric).

Northlane operating system

The dashboard that reads the inbox, tracks every RFQ and drafts the quotation.

A web application that I directed Claude Code to build. Every day it reads both mailboxes, drops the noise, sorts each thread, matches supplier replies to the customer request and proposes the next action. Two quotation routes pick the winning supplier for each line and price the landed cost step by step. A margin check learns from past quotes and flags anything that looks too thin.

What came of it293 supplier RFQs issued through it (October 2026).

Estimating brain and price memory

Eight years of supplier emails turned into a priced, sourced memory the agents estimate from.

I had the agents read supplier emails and attachments back to 2018 and build a searchable price record, a product catalogue and a pricebook where every rate carries its source and date. On top sits an estimating method: every number carries an estimate class, a likely and a cautious value, and a basis date, and contingency comes from a risk simulation that went through an independent audit.

What came of it2,356 priced records from 137 suppliers, a 5,112-product catalogue, a 236-row pricebook, and 13 of 13 audit findings accepted.

Arabic tender reader and Etimad feed

Turns a scanned Arabic tender into structured line items, and watches the government tender portal daily.

A single pass over a 99-page scanned Arabic tender agreed on only 30% of part numbers, so I changed the design: render each page at 600 dpi, run independent readers, and let them vote. The same habit sits behind a daily feed that scrapes open and awarded Saudi government tenders, matches Arabic titles to our supplier database and posts a digest to Telegram.

What came of it4,362 structured line items from one 99-page scan, with independent readers voting.

Drawing-to-costing system for an engineering group

My first outside client: estimators drop drawings in a folder and get a reviewed estimate back.

Built for a 400-person Saudi engineering group that came to me by referral, running on their own on-premise server. Watched folders pick up new drawings, a reader per discipline extracts quantities, and an hourly email goes to the estimators on the group's own domain with prior rates and where each came from. Every hand-off is logged. Seventeen estimators use it daily.

What came of itDrawings to a reviewed estimate went from two to three weeks down to about one day (client-reported; described anonymously with the client's agreement).

ProCul, the agent-native ledger

An accounting product for Saudi companies where agents post, match and prepare the close, and a person approves every entry.

The kind of AI-native general ledger Rillet is building in the US, shaped for Saudi companies. ProCul mirrors a company's books into its own double-entry ledger, matches payment settlements and bank lines to the invoices behind them, drafts the entries, prepares the VAT return and assembles the monthly close. The arithmetic is plain code with money held as whole halalas, and a model only reads documents and suggests categories. Nothing posts on its own: every entry waits as a draft until an accountant approves it. I directed Claude Code to build it.

What came of itTested on a synthetic Saudi company with 20 planted errors: it found the exact SAR 1,677,448 difference and flagged exactly the twelve account slots they touched. 1,238 automated tests pass. No customers yet.

@heyaliasif content engine

A daily video pipeline: raw clip in, reviewed and captioned edit out, posted only after I approve.

I hand the pipeline a clip. It transcribes offline, cuts, captions, renders in English and Arabic, and runs the result past a multi-lens review before I see it. Claude drafts the concepts and checks each one against a red-team pass for stale news and claims I can't source. Posting goes through official schedulers and APIs and I press send myself.

Chief-of-staff morning brief

A bounded agent that reads every project each morning and hands me a short decision brief.

Each morning it checks what changed across my projects, sorts the loose notes I dropped during the day, drafts what it can and writes one brief with the decisions waiting for me. It never sends anything on its own. A nightly pass turns each day's work into a shared log, so every project knows what the others learned.

The AI radar on this site

A publishing pipeline that turns primary AI sources into a short post with my take, up to three times a day.

A fetcher pulls lab blogs, trending models and daily papers from Hugging Face, GitHub releases, Reddit, Hacker News and a handful of YouTube channels, then drops anything already seen. Claude writes a post only if at least three items clear a value bar, and every item must link the primary artifact: a model card, repo, lab post or paper. Links are checked, the copy is scanned, and the post goes live with a note to me.

Saudi Tender Tagger

A small model that reads an Arabic Saudi government tender title and says what kind of work it is.

I trained it on more than 7,000 public Etimad tenders. On tenders it never saw, it ranks the right one of 54 activity labels first 61% of the time and in its top three 80% of the time, and it gets the procurement type right 88% of the time. It is for screening a long list with a person in the loop, not for deciding what to bid on. You can try it on this site.

Try itModelDatasetCode