I'm 74 and Built an AI Content Engine That Uses 5 Competing Models as Quality Control. Here's What 161 Posts Taught Me
Hey IH 👋
I'm Mohan. 74 years old. Industrial engineer for 50 years. WordPress developer for 12. AI builder for the last two.
Last month I published 161 blog posts in 48 hours. Average SEO score: 84/100 on Rank Math. Not with a team or agency — with a methodology I built myself.
The problem
AI-generated content has a trust problem. Not because it sounds bad — it sounds great. That's the danger.
I fact-checked an early GPT-4 draft. Three claims. Confidently stated. All wrong. As an engineer, that's unacceptable.
What I built
Pentangulat — a 5-AI consensus validation system.
Every factual claim runs through GPT-4, Claude, Gemini, Mistral, and Cohere. The rule: 70% consensus. If fewer than 4 out of 5 models agree on a claim, it gets flagged.
Result across 161 posts: ~30% of initial drafts had at least one claim that failed consensus. That's nearly one in three posts that would have gone live with errors.
Tech stack (for the builders)
- Python 3.11
- httpx (no SDK dependencies)
- SQLite (zero infrastructure)
- Anthropic API for content generation
- 3,536 lines of production code
- 111 passing tests
The product
I documented the entire methodology in the AI-Era SEO Playbook — covering GEO (Generative Engine Optimisation), LLMO (Large Language Model Optimisation), AEO (Answer Engine Optimisation), Rank Math configuration, and a 90-day rollout plan.
$47.65 on Gumroad → seekrates-ai.com/playbook
Why $47.65? Because one SEO consultant's hourly rate buys you a permanent methodology.
Results so far
- 161 posts live on seekrates-ai.com
- 84 average Rank Math score (industry typical: 60-70)
- ~30% error catch rate via consensus
- Schema markup (FAQ + Article) on every page
- llms.txt + AI crawler configuration live
- All sitemaps submitted to Google + Bing
Lessons for other indie hackers
-
Age is irrelevant. Curiosity is the only prerequisite. I started building AI tools at 72.
-
AI consensus is massively under-explored. Everyone uses AI to generate. Almost nobody uses AI to validate. That gap is where quality lives.
-
Boring infrastructure first. I spent weeks on vocabulary definitions and test infrastructure before writing a single feature. The code worked because the requirements were unambiguous.
-
Walk the talk. 161 live posts with verifiable scores. Not a landing page with promises — a live site with proof.
Would love to hear from anyone working on AI content quality, SEO automation, or multi-model validation. Happy to share technical details.
What's your approach to AI content accuracy?
The 5-model consensus approach is clever — it's essentially using disagreement as a signal for uncertainty. Most AI tooling treats models as interchangeable; you're treating them as a panel of experts with different biases.
A few questions on the methodology:
Consensus failure modes: When 4/5 models agree but they're all wrong (shared training data bias), how do you catch that? Do you have a fallback for factual claims in domains where all models tend to hallucinate together (e.g., recent events, niche technical specs)?
Cost per post: Running 5 API calls per claim adds up. What's the actual per-post cost for a typical 1,500-word article? Curious if this is economically viable at scale.
Threshold tuning: Why 70% (4/5) specifically? Did you experiment with stricter thresholds (5/5 required) or looser ones (3/5)?
The "boring infrastructure first" lesson resonates. I've seen too many AI projects skip vocabulary alignment and end up debugging semantic mismatches for months.
Also — starting AI development at 72 is genuinely inspiring. The curiosity-over-credentials mindset is underrated in tech.