← Projects
Personal project 200+ practitioners Live

AI Intelligence Briefing

A weekly AI digest for supply chain and procurement practitioners — built, hardened, and shipped without me in the loop.

The problem

Keeping up with AI in supply chain and procurement meant hours of reading every week — newsletters, research, announcements — with no reliable way to separate signal from noise. Worse, dozens of colleagues were each doing that same reading privately, with no shared view of what mattered. The shortage was never information. It was someone to sift it down to what's actually worth knowing.

What I built

Every week, a fully automated "editor" does that sifting. It reads the week's AI news from across the industry, judges what actually matters to supply chain and procurement, writes a concise strategic brief in a consistent house style, and emails it — designed and formatted — to more than 200 colleagues. Nobody has to write it, and nobody needs to be in the loop for it to go out.

Behind that simple result is a multi-stage pipeline I took from prototype to leadership sign-off to production. It gathers candidate articles from nine industry newsletters and seven semantic searches, then scores each one for relevance with logic tuned to the field — weighting headlines, fading older stories, and capping any single buzzword so keyword-stuffed pieces can't game their way to the top. Around thirty summaries are written in parallel by a fast, inexpensive model (Claude Haiku), and a single pass from a stronger model (Claude Sonnet) turns them into the finished brief. It then renders itself as a branded email and PDF and sends them. To stay fresh, it remembers its last eight issues and is told to find a new angle rather than repeat the previous week's themes.

Tech stack

Python Claude Haiku 4.5 Claude Sonnet 4.6 Exa Search Jinja2 WeasyPrint feedparser Pydantic tenacity Gmail SMTP

Architecture

How a week of scattered reading becomes one finished brief, left to right — gather, score, summarise, write, deliver.

Data Sources
Substack RSS 9 feeds
Exa semantic search 7 queries
Scoring
Keyword scoring recency decay, 3-hit cap per keyword
Deduplication URL + theme-level
AI Processing
Haiku summarise ×~30, parallel
Pydantic validation with retry logic
Editorial
Sonnet strategic brief 8-issue rolling history context
Delivery
Jinja2 → WeasyPrint PDF
Gmail SMTP
history.json history updated

The hard part

The hard part was giving it editorial judgement without an editor. Stopping it from running the exact same article twice is easy. The real risk is subtler: the same themes resurface week after week in different stories, and a brief that keeps making the same point gets ignored. So each week the system is shown what the last issue actually said — its main theme and how it opened — and told to find a fresh angle, while still allowing a deliberate callback when this week genuinely follows on from last.

The scoring needed the same kind of care. Capping how much any one keyword can count — no more than three hits per article — stops pieces that simply repeat fashionable terms from crowding out the ones with something real to say.

Screenshots

AI Intelligence Briefing — newsletter preview AI Intelligence Briefing — pipeline run

Outcomes

  • Delivered every week to 200+ supply chain and procurement practitioners
  • Nominated for a firm-wide Innovation Award (2026); approved by practice leadership
  • Runs start to finish in three to five minutes, with nobody in the loop
  • Remembers its last eight issues to avoid repeating itself, and caches its work so re-running it costs nothing