Post-mortem · October 2026
From August to October 2026 I built and operated Newsig — a financial-news pipeline that ingested 23 sources, deduplicated stories into events, translated them, mapped each to the assets it moved, and served it all as a server-rendered site. I wrote almost none of the code myself, in the sense that I typed it. The interesting part is not that it ran. It is where the line between me and the machine actually sits.
Roughly 90% of the implementation, and effectively 100% of the day-to-day operation. Concretely:
36,354 lines across a Preact frontend, a Node/Express/SQLite backend, 25 operational scripts and a migration layer. Server-rendered pages, an LLM translation pipeline, an asset-matching engine, sitemap and IndexNow submission, an RSS feed, and an editor-facing admin surface. I specified behaviour and reviewed output; the machine produced the code.
A 283-line idempotent bootstrap that provisions users, SSH hardening, firewall, Node, pm2, nginx and cron from a blank Ubuntu box — and an 80-line deploy script with automatic dist rollback and a pre-deploy database snapshot. Deploys were one command, and a failed build restored the previous version without downtime.
Five cron jobs: a health check every five minutes, a backlog monitor every thirty, a daily traffic report, a daily digest push, and a nightly hot backup. pm2 restarted the process 286 times over 69 days — crashes were absorbed without me looking. That is not a claim of perfection; it is a claim that nobody had to be awake.
When I suspected the traffic spike was fake, I described the suspicion and the machine wrote the queries, pulled the raw hits, compared device fingerprints across countries, and produced the evidence. That loop — hypothesise in plain language, get a measurement back — turned out to be the most valuable capability of all.
The fourth item in the right column is the one that killed the project. I spent four months on the product and roughly zero days on channels — no social account, no newsletter, no audience. In October the site received 27 page views and had zero indexed pages. A product with no distribution is a private experiment that happens to have a URL.
The build was about 90% delegated. The maintenance was about 100% automated. The judgement was 0% delegable — and that is exactly why it still mattered.
I shipped too much: sentiment pages, watchlists, accounts. None of it moved the needle. A two-week version in front of ten people would have taught me more than four months of building.
But the capability is real and it transfers. Knowing that one person can take a system from blank server to unattended production — code, deploy, monitor, back up, diagnose — changes what is worth attempting. The next projects will start from that assumption rather than treating infrastructure as the hard part.
The failure was not technical and it was not about the model. It was that I optimised the half of the work I found interesting, and skipped the half I found uncomfortable.
The site is frozen as a static snapshot: 489 pages, no JavaScript, every story readable as plain HTML, with the original URLs intact. It costs nothing to keep and it is an honest record of what four months of solo work looks like — including the parts that did not work.