We Published 20+ AI-Generated Research Articles in 30 Days — The System and the 3 Things That Broke
An autonomous content engine published 20+ research articles in 30 days. Here's how the system works — and the 3 things that broke along the way.
We published 20+ AI-generated research articles in 30 days. Here’s the system — and the 3 things that broke.
Since late March 2026, Charaka Notes — our public research publication — has been publishing data-driven analysis five days a week. Over 20 pieces covering startup survival patterns, sector intelligence, funding dynamics, and AI-native operations.
Not a human typing articles. An autonomous content engine.
How the system works:
The engine pulls patterns from our knowledge graph (100,000+ entities, 14,000+ companies, 365 postmortems). It identifies what’s interesting, drafts analysis, runs fact-checking against the KG, and publishes. A human approves the final output — never the process.
Same principle as the “dark factory” in software engineering: specifications drive the work, agents execute, humans approve outcomes.
The output quality isn’t “AI-generated blog post” quality. These are structured research pieces with real data, specific company references, sector-level pattern analysis, and actionable investment implications. Because the knowledge graph is the source of truth, hallucination risk drops dramatically.
The 3 things that broke:
1. The deployment pipeline. Our Cloudflare token expired silently. The content engine generated and approved three articles that never reached the website. No error. No alert. We only noticed when someone asked why Tuesday’s piece was missing. Fix: health checks that verify the published URL returns 200 after every deploy.
2. The quality gradient. Early pieces were consistently good because the topics were obvious — clear patterns, strong data. By week three, the engine was reaching for thinner signals. Pieces started getting vague. Fix: a minimum-evidence threshold — if the KG doesn’t have at least three independent data points supporting a claim, the claim gets cut. Quality dropped before we noticed. Lesson: autonomous systems need quality floors, not just quality ceilings.
3. The reader feedback loop. We published an analysis of multi-agent architecture — why the best firms will run 100 agents per partner. An ESSEC-trained analyst read it and came back with five structural gaps: agent scaling logic was asserted without framework, coordination architecture was invisible, bias prevention was absent. Detailed, specific, devastating feedback. We rewrote the entire piece. Then he came back again with the antifragility question — what happens to human judgment when agents handle 95% of analytical work? We added that too. Fix: the engine now produces content, but the calibration comes from readers. The best quality signal isn’t internal review — it’s an informed reader who cares enough to tell you what’s missing.
The meta-lesson:
Building an autonomous content engine is not primarily a writing problem. It’s a systems engineering problem. Token management, deployment pipelines, quality thresholds, audience feedback loops — the same calibration discipline we apply to investment analysis applies here.
The engine is still running. Still learning. Still getting better.
Read Charaka Notes: https://getmanthan.com/charaka-notes/
Read more at getmanthan.com
Mayank Mathur | Founder, Manthan Intelligence | GP, Tavaga Fund
Never miss an insight
Free dispatches, every day. Unsubscribe anytime.
No spam. Just intelligence.