DEV Community: Олександр

Making AI-written content sound less like, well, AI

Олександр — Tue, 14 Jul 2026 04:20:08 +0000

Hey everyone,

I wanted to share a small technical detail from working on my content sites. When I started integrating AI-assisted writing, the biggest challenge wasn't generating content, but making it sound human. Initial drafts often felt flat, predictable, and frankly, a bit robotic.

My solution involved a few post-processing steps. First, a 'burstiness' pass. This isn't about randomizing, but analyzing sentence length variance and adjusting it to mimic human writing patterns, which are rarely uniform. I found targeting a standard deviation of 4-6 words in sentence length within a paragraph usually did the trick.

Then came a pass specifically for contractions. AI models often avoid them, making text sound overly formal. I built a simple replacement engine that identifies common opportunities for contractions (e.g., 'it is' to 'it's', 'they are' to 'they're') and applies them where grammatically appropriate. This significantly increased readability.

Another small thing was injecting specific numbers. AI tends to be vague. Instead of 'a lot of people,' I'd try to find a way to insert 'over 200,000 people.' It just feels more real.

Finally, a 'banned words' sweep. I maintain an evolving list of about 80 words and phrases that, to me, signal AI generation (e.g., 'delve into,' 'unlock the potential,' 'tapestry of'). The system flags these for manual review or suggests alternatives. You can see the effect of these processes on content like the weekly awards on aceju.com or the brewing guides on dreqo.com.

Happy to answer any questions about the specifics of these post-processing steps.

daily qa pipeline

Олександр — Tue, 07 Jul 2026 01:36:16 +0000

I've been working on a daily QA pipeline for my content-site network, which scores every article on depth, structure and anti-AI patterns. The pipeline is designed to ensure that our articles are informative, well-structured and not easily replicable by AI tools. For example, on fumpe.com we have a large collection of practical pet care guides, and this pipeline helps us maintain their quality. The pipeline runs daily, automatically regenerating articles that score low on these metrics. I've also applied this pipeline to dreqo.com, where we have a wide range of coffee brewing and gear guides. The results have been promising, with a significant improvement in article quality across the network. The pipeline uses a combination of natural language processing and machine learning algorithms to evaluate each article. If you're interested in the technical details of how this pipeline works, I'd be happy to answer any questions about it.

I built an LLM router that picks a different model per article type

Олександр — Sun, 28 Jun 2026 01:37:26 +0000

I've been quietly running a small network of niche content sites for a few weeks. Two are live on real domains:

afixu.com - home improvement and DIY tool guides
aceju.com - weekly AI tool awards, editorial format scored across 4 criteria

The niche isn't the interesting part. The pipeline is.

The LLM router

Instead of one model for everything, a router picks per article type:

premium/commercial pieces go to a stronger model
high-volume informational drafts go to a cheaper, faster one
if a provider is rate-limited, it falls through a cascade so generation never stalls

The QA loop

Every draft, before it ships, passes through:

an anti-AI-tells cleanup (80+ regex replacements - "delve", "leverage", "robust", etc.)
a rankability judge scoring depth, specificity, intent-match and trust on an 8-point scale
a structure/schema check (FAQ block, JSON-LD, internal links)

A daily job re-scores everything published in the last 24h and auto-regenerates the failures. The whole thing is Next.js 16 + MDX, one git repo per site.

The honest target: a $30k/mo run-rate in ~24 months across 30 mature sites, monetized through guest-link sales, not ads. Slow on purpose.

Happy to go deeper on the router logic or the QA scoring if anyone's interested.