Building an AI Curation Assistant for Steem: A Developer's Log — Part 1

The Problem I Want to Solve,
Anyone who has spent time browsing Steem's trending or new posts knows the challenge: genuinely good content — well-researched, original, thoughtfully written — often gets buried under low-effort or spun posts, simply because there's no easy way to separate the two at scale. Curators do their best, but manual review doesn't scale to thousands of daily posts across hundreds of communities.

That's the problem I want to chip away at: a lightweight AI-assisted tool that helps flag posts worth a closer look, without replacing human judgment.

What I'm Building

The project is a Python-based bot — I'm calling it a Curation Assistant for now — that combines two things:

1. Steem blockchain data, pulled via the beem library, covering post metadata: author history, post length, tags used, and engagement patterns.
2. Basic text-quality signals, using simple natural-language processing (not a full LLM at this stage) to flag characteristics like unusually repetitive phrasing, very short body text relative to claimed effort, or content that closely mirrors another recent post on the chain — a lightweight, local plagiarism check.

The output isn't a verdict. It's a short, scored summary — something a curator or a community moderator could glance at before deciding whether a post deserves a deeper read.

Why "AI Machine" Undersells It (For Now)

I want to be upfront about where this project actually stands: this is not a trained machine-learning model yet. Right now it's a rules-based scoring script — pulling on-chain data and running straightforward text heuristics. Calling it "AI" at this stage would be overselling it.

The roadmap toward something more genuinely AI-driven looks like this:

• Phase 1 (current): Rules-based scoring using on-chain metadata + simple text statistics.
• Phase 2: Introduce a small classification model (likely a fine-tuned lightweight transformer, run locally) trained on a labeled sample of high-quality vs. low-effort Steem posts.
• Phase 3: Feedback loop — allow curators to confirm or reject the tool's flags, and use that signal to improve scoring over time.

I'd rather build in public and be honest about which phase I'm in than dress up a scoring script as "AI" prematurely.

Technical Stack (So Far)

• Language: Python
• Blockchain access: beem library for Steem API calls
• Text processing: basic NLP with standard Python libraries for the current heuristic phase
• Data storage: local JSON for now, likely migrating to a lightweight database once the account sample size grows

What I Need Help With

This is very much a first-iteration build, and I'd welcome input from anyone else working on Steem tooling:

• Has anyone already built something similar for content curation on Steem? I'd rather build on existing work than duplicate it.
• Suggestions on which on-chain signals are most predictive of post quality, from people who've curated for a while.
• Any pitfalls with beem's rate limits I should plan around before running this against larger post samples.

Next Steps

My next post in this series will share the first working version of the scoring script, along with real output on a small sample of recent posts — not simulated numbers, actual results I can show and let people question.

---

This is an original, in-progress development project. No code or output has been fabricated — the roadmap above reflects the actual current state of the build.