The Telegram channel X-ray that says more about AI content tooling than about Telegram
If you run social accounts for a living, you already know the dirty secret of “AI-powered analytics”: most of it is a chat model wearing a dashboard. You paste a year of posts, you ask for a summary, you get a confident paragraph that’s 60% vibes and 40% hallucination, and you have no way to audit a single claim. So when a maker shows up on Product Hunt with a tool that explicitly refuses to write text — that returns typed probabilities instead of prose, in roughly 300 ms per post — that’s not a novelty. That’s a different architectural bet about what AI is for. And for anyone whose job is deciding what to publish across Instagram, TikTok, YouTube, X, LinkedIn, Facebook, Threads, and Pinterest, that bet is worth understanding, because it’s the same bet you should be making about your own content ops.
What Jev Wrapped actually does, minus the launch-page gloss
The product is Jev Wrapped, built by Ivan Gabor on top of Jev, the decision model from TypeSafe. The pitch is narrow and specific: point it at a Telegram channel, and it produces an “X-ray” of that channel’s content mix. Not a summary. A classification.
Here’s the mechanic, per the maker’s own description: Jev takes a post plus four typed questions and returns a probability for each answer, in about 300 ms. No generated text at all. The output is a set of shares across roughly ten content kinds — news, opinion, ads, jokes, the author’s own promotion, and others — plus three separate percentage scores for ads, clickbait, and emotional pressure, plus a month-by-month mix for the last twelve months.
The maker’s own demo number: analyzing Durov’s channel took 6 seconds and cost $0.0043, and came back with 53% of posts being about his own projects. That’s the entire value proposition in one line — a year of channel history, read post by post, for less than half a cent. The cost and latency figures are the maker’s claims, not independently verified, but they’re the kind of claim that’s easy to falsify, which is itself a good sign.
It’s open source under MIT, runs as a single Cloudflare Worker on the free plan, and works with your own TypeSafe or OpenRouter key. You can try it at wrapped.ivanhabor.com.
The part most coverage will skip: it admits when it’s wrong
Buried in the launch copy is the most operationally important sentence: “Jev makes mistakes. On a channel with no paid advertising it still marked two partner promotions as ads.”
That’s a false positive, disclosed by the maker, unprompted. And the mitigation is the interesting design choice — below every card, the page lists the posts it scored highest, with links, so you can check the model’s work yourself. In my experience testing classification tools for content audits, this is rare. Most vendors hide the per-item scores and show you an aggregate, because aggregates are unfalsifiable. Showing the top-scoring items is how you turn a black box into something a skeptical operator can actually audit in five minutes.
Why this matters more than Telegram
Telegram is not most creators’ primary distribution channel. It’s a newsletter-ish, community-ish surface that matters a lot for crypto, news, and certain creator niches, and barely at all for a TikTok-first fashion creator. So on its face, a Telegram channel X-ray is a niche tool for a niche audience.
But the pattern underneath it is the thing. Let me explain what I mean, because this is where I think the launch is more instructive than the product.
The three failures of “summarize my content with AI”
When I’ve run content audits — the kind where you pull 12 months of posts and try to figure out what actually worked — I keep hitting the same three walls with chat-model-based tooling:
First, summaries destroy the signal you need. If you ask a model to summarize a year of posts, it will tell you themes. It will not tell you that 34% of your output was repackaged promotional material, or that your “educational” posts were actually clickbait with a lesson stapled on. Themes are for strategy decks. Proportions are for decisions. You can’t fix a content mix you can’t measure.
Second, prose output is unauditable. When a model writes a paragraph, you can’t check its work without reading everything yourself — which defeats the purpose. When a model returns “this post: 0.91 ad, 0.04 news, 0.03 joke,” you can spot-check instantly. Probability-per-question is a fundamentally more honest output format than a summary, even when the probabilities are wrong, because wrong probabilities are visibly wrong.
Third, per-item cost and latency determine what’s even possible. This is the part operators underestimate. If each post costs a meaningful fraction of a cent and takes seconds, you can only sample. If it costs $0.0043 for an entire channel and runs in 6 seconds, you can read everything. The maker’s framing — “at that speed and price you can read a whole year of a channel post by post instead of asking a chat model for a summary” — is the real thesis. Full-corpus analysis beats sampled summarization every time, and it’s only recently that the economics allowed it.
Why TikTok creators should care more than LinkedIn ones
Here’s my take on where this pattern lands unevenly across platforms.
TikTok and YouTube Shorts creators live and die by content mix in a way LinkedIn creators mostly don’t. On TikTok, the algorithm rewards consistency of signal — if your account is legibly “one thing,” distribution is easier. If you’re flip-flopping between trend-chasing, personal vlogs, and product promos, the system struggles to place you, and your watch time and completion rates get noisy. A tool that tells you “38% of your last 90 posts were ads, 12% were jokes, 4% were actual opinion” is directly actionable: you can see when you drifted off-brand.
LinkedIn is different. The LinkedIn algorithm rewards a narrower band of formats, and “personal promotion masked as a lesson” is basically the native genre. Knowing you’re 60% self-promotional on LinkedIn isn’t a correction — it’s a description of the platform. So the diagnostic value is lower.
Where it gets interesting for everyone is emotional pressure and clickbait scoring. Those two percentages are the ones I’d actually pay for. Most creators genuinely don’t know how much of their output is engineered outrage or manufactured urgency, because it happens gradually. A monthly trend line of “your emotional-pressure score went from 11% to 29% over six months” is the kind of feedback that changes behavior.
What social teams can steal from this, even without Telegram
You don’t need to run Jev Wrapped to apply its logic. Here’s what I’d actually borrow, and I’ve started doing versions of this in my own workflow.
Classify, don’t summarize. When you do a quarterly content review, stop asking an LLM to summarize. Ask it to classify each post against a fixed taxonomy — educational / opinion / promo / community / entertainment / news — and return counts. You’ll get a content mix you can act on. This is the same principle behind Buffer’s analytics and Metricool’s reporting, except those tools measure performance metrics, not content intent. Content intent is a separate axis, and most stacks don’t measure it at all.
Demand per-item auditability. Any AI tool you adopt for content work should let you see the items behind the score. If it only shows you an aggregate, you can’t trust it, and you can’t catch its failure modes. Hootsuite, Later, and most scheduling suites are fine at this for scheduling; for AI scoring, hold the bar higher.
Run the math on full-corpus analysis. If your tooling budget is $50/month and per-item analysis costs fractions of a cent, you can afford to analyze everything you’ve ever published, not a sample. The reason most teams sample is historical cost, not necessity. That constraint has moved, and most content ops haven’t updated.
Where the math breaks
I want to be careful here, because the maker’s numbers are seductive and I haven’t independently verified them. The $0.0043 figure is for one channel — Durov’s — which is a specific length and posting cadence. Not disclosed: how cost scales with channel size, whether very long posts blow past the model’s context, and whether the 300 ms figure holds under load or is a best-case local benchmark. Those are open questions, and any operator planning to run this across 50 client channels should treat the cost model as unproven until they test it themselves.
There’s also a subtler issue: classification accuracy on satire, irony, and mixed-intent posts. A post that’s simultaneously a joke and an ad — which is a huge fraction of creator content — is exactly where a four-question typology will struggle. The maker’s disclosed false positive (partner promos flagged as ads on an ad-free channel) is a hint that the model’s boundaries between “promotion” and “partnership” and “recommendation” are fuzzy. That’s not a fatal flaw; it’s a known limitation you’d need to calibrate against your own content before trusting the percentages.
Where my judgment says this falls short
Let me be direct about the product’s actual limits, because the launch copy is good and I don’t want to just echo it.
It’s Telegram-only, and Telegram isn’t where most creator revenue lives. If you’re a YouTube creator or an Instagram operator, this tool does nothing for you today. The maker hasn’t announced expansion to other platforms, and I’d bet the classification model would need retraining per-platform anyway, because what counts as “clickbait” on X is different from what counts as clickbait on Pinterest.
It’s a diagnostic, not a prescription. Knowing your content mix is 53% self-promotion doesn’t tell you what to do about it. The tool tells you what you published; it doesn’t tell you what performed or what to publish next. Pairing it with real performance data — UTM tracking on outbound links, platform-native analytics, engagement rate by post — is on you.
The open-source, bring-your-own-key model cuts both ways. It’s cheap and transparent, which is great. It also means you’re responsible for your own API keys, rate limits, and reliability. If your OpenRouter key hits a rate limit mid-audit, that’s your problem, not the maker’s. For a solo operator, fine. For an agency running this across dozens of clients, you’d want to wrap it in your own infrastructure.
Who it’s NOT for: anyone whose primary channels are Instagram, TikTok, YouTube, X, LinkedIn, Facebook, Threads, or Pinterest. Anyone who needs scheduling, publishing, or analytics in the same tool. Anyone who wants a done-for-you SaaS with support and SLAs. This is a sharp, narrow instrument for a specific job — auditing Telegram channels — and it’s honest about that.
What I’d watch / test next
Three concrete things I’d do this week if this pattern interests you, in order of effort.
One: run the tool on a Telegram channel you actually know. Pick a channel whose content you understand well — maybe a competitor’s, maybe your own — and check the maker’s top-scored posts against your own judgment. That’s the fastest way to calibrate whether the classification matches your read of the content. If it does, the pattern is worth stealing; if it doesn’t, you’ve learned something about the limits of probability-based classification. The maker explicitly asked for this: “If a card looks wrong for a channel you know, tell me here. I want to see where the model misses.”
Two: build a poor-man’s version for your main platform. Take your last 90 posts, define a 6-category taxonomy, and run them through your LLM of choice with a strict “return counts only, no prose” instruction. You’ll spend maybe a dollar and get a content-mix baseline you’ve never had. This is the single highest-leverage thing most creators could do this quarter, and almost nobody does it.
Three: watch whether the pattern spreads. If classification-not-summarization becomes the default mode for content AI — and I think it will, because it’s cheaper, faster, and auditable — expect the incumbents to bolt it on. Canva, CapCut, and the scheduling suites all have the distribution to ship this as a feature. The question is whether they’ll ship it with per-item auditability or hide it behind an aggregate score. That’s the detail that separates a useful tool from a plausible-sounding one, and it’s the detail I’ll be watching.
