I was testing AI models. Then a newsroom asked for a demo of CoTel.
I thought my biggest problem right now was figuring out LLM economics.
Over the past few days, that’s exactly what I’ve been working on:
testing 8 models from different providers, comparing quality, speed, and cost, and trying to understand how AI credits and pricing should actually work for CoTel.
I already have some interesting conclusions.
For example, Claude Haiku performed worse than comparable OpenAI and Gemini models on several simple tests while costing noticeably more. So I removed it from consideration completely.
But at the same time, something much more important happened.
The journalist I mentioned in one of my previous posts told his manager about CoTel.
And suddenly, the head of their department wanted to see a live demo.
That honestly surprised me a lot.
Because they actually have a real unresolved problem around Telegram workflows.
They monitor many local news chats and channels where videos, photos, incident reports, links, and other important signals appear very quickly.
Sometimes messages get deleted by admins.
Sometimes information simply disappears inside huge streams of messages.
And during that conversation, I suddenly realized something important:
for their workflow, the most valuable feature may not even be chat Q&A itself, but the subscription system I already built into the MVP.
Because subscriptions can periodically monitor chats automatically.
So instead of:
“summarize this chat”
the workflow becomes:
“every 5 minutes, scan these chats and immediately send me all new videos, images, or other relevant signals.”
Which means journalists no longer need to sit inside Telegram all day manually monitoring dozens of sources.
Right now I need to improve subscriptions so one subscription can monitor an entire group of chats instead of just a single one.
And honestly, what motivates me most is the fact that for the first time this is no longer just:
“interesting project.”
Now there’s a real person saying:
“If this works properly, we’re ready to pay for it.”
That’s a very strong signal for me.
Until now, CoTel still felt like my personal experiment and first MVP that I’ve been building at night during maternity leave.
But now I’m starting to see real workflows where the product can genuinely save people time and solve an actual operational problem.
And that changes how the whole project feels.
Now I’m curious whether this kind of “AI monitoring workflow” exists in other industries too, not just journalism.
And have you ever discovered that your product’s first real paying use case was completely different from the one you originally imagined?
That's a wild pivot - went from casually testing models to a newsroom reaching out. Did they come to you, or did you reach out to them first?
The story is actually pretty simple.
A good friend of mine, who works as a journalist, was testing CoTel for me and sharing feedback based on his real work workflows. He found the product interesting and mentioned it to his manager during one of their newsroom meetings.
That's how the editor of a European newspaper ended up reaching out to me.
If you're curious, I wrote more about this in the post - https://www.indiehackers.com/post/i-built-a-telegram-ai-tool-for-myself-users-immediately-wanted-different-things-7e4a13a410
The newsroom angle is interesting because newsrooms get hit harder by AI accuracy issues than most categories - getting facts wrong is existential, not just a brand problem. The pattern I keep running into when testing AI models is platform variance. The same query to ChatGPT, Perplexity, Gemini, and Claude returns meaningfully different answers, and the disagreement rate matters more than any single platform's score. In a 244-company benchmark across all four, the most-cited entity often had the worst accuracy on a given platform. Curious what your newsroom contact is actually testing CoTel against - model-vs-model disagreement, accuracy against ground truth, or both?
That's an interesting observation. By the way, I haven't included Perplexity in my testing yet — currently I'm only benchmarking models from OpenAI, Google, and Anthropic. I'd be curious to hear your opinion: how good are Perplexity's models, and which models would you personally rank at the top right now?
As for the newsroom, I think they'll evaluate CoTel primarily on how well it solves their specific workflow: monitoring large numbers of local Telegram chats and quickly detecting newly posted media files.
I'm also gradually coming to the conclusion that most users don't actually care which model is being used. They care about the outcome. That's why I'm testing different models across different query types and plan to build a routing layer that automatically selects the most suitable model for each task. All of that will stay under the hood — users will only care about how effectively the system solves their problem.
Inbound from a newsroom while you're still figuring out LLM pricing is a great problem to have. That's the kind of pull that tells you the product is on to something real.
The LLM economics question is one a lot of AI founders are wrestling with right now. What you spend on inference vs. what the customer actually values rarely lines up cleanly. The pricing model often has to get creative to bridge that gap.
Curious how the newsroom demo went and whether it turned into something.
Yes, that's exactly where I am right now.
I'm essentially building my own routing logic between different models, analysis depths, and usage scenarios.
The challenge is that costs and quality requirements can vary dramatically depending on what the user is actually trying to accomplish.
I suspect most AI products eventually have to go through this stage.
I'm actually preparing a new post about the results of testing around 10 models from different providers, and I've already discovered some pretty interesting patterns.
The newsroom demo will happen a bit later. First, I want to finish the LLM router and implement a few improvements that came directly from their feedback.
Thanks for your interest in the project!
This is a very strong signal because the use case moved from “AI chat over Telegram” to operational monitoring of live information streams.
For journalists, the pain is not just summarization. It is not missing important signals while dozens of chats keep moving, posts get deleted, media disappears, and the team has to react quickly. That is a much more valuable workflow than Q&A.
I’d seriously lean into that direction: AI monitoring for fast-moving Telegram/source channels. Journalism is the first wedge, but the same pattern could apply to security teams, local operators, logistics, investigations, crisis monitoring, market research, and any team watching noisy information streams.
One thing I’d pressure-test now is the name. CoTel is short, but it does not immediately carry the “signal monitoring / intelligence layer” idea you just discovered. Since a real department is already asking for a demo, the brand may need to feel more serious than a personal MVP before more buyers see it.
Exirra .com would fit this direction better as a broader signal-intelligence brand. It keeps the product on your side: same Telegram workflow, same MVP, but with a name that can grow into AI monitoring, alerts, subscriptions, and operational intelligence beyond the first newsroom use case.
This feels like one of those moments where the use case became clearer than the original product idea. The name should probably follow that stronger use case before demos and early buyer memory start locking around CoTel.
Actually, I had been thinking about the subscription mechanism from the very beginning, and it’s already implemented in the MVP. Right now the main missing piece is the ability for a single subscription to monitor an entire group of chats, and most likely that’s what I’ll work on next.
But what really surprised me was how useful this workflow could become specifically for teams and professional workflows.
Because my own original use cases were much more personal and narrow:
tracking the sale of a specific item,
following apartment rental posts,
getting summaries from chats I care about so I don’t spend hours inside Telegram.
I wasn’t looking at it as a full operational tool for professional teams.
So it’s genuinely inspiring to see that the same mechanism may actually help people who use Telegram as part of their daily work and need to react to information very quickly.
As for the naming discussion — I’m still not sure I want to move toward a more abstract brand name.
For me, the logic behind CoTel is still very straightforward:
Copilot for Telegram.
And honestly, names like Exirra currently feel more like random letter combinations to me without an immediately clear association or meaning behind them. Maybe I’m just missing the context of how exactly you arrived at that name 🙂
That’s fair, and I wouldn’t dismiss CoTel for the current MVP. “Copilot for Telegram” is clear, especially while the product is still centered around personal Telegram workflows.
The reason I brought up Exirra is because the use case you’re describing now feels less like a Telegram companion and more like a signal-monitoring layer.
That shift changes the naming requirement.
CoTel explains the starting mechanism. Exirra would be more of a category brand: abstract, serious, and broad enough to cover monitoring, alerts, subscriptions, source tracking, and operational intelligence without being locked to Telegram forever.
Invented names can feel random at first, but they work when the product needs to own a category instead of describe a feature. Stripe, Linear, Vercel, Datadog, Sentry, Palantir — the meaning came from the product, not the literal word.
So I’d frame it this way:
If CoTel stays “Copilot for Telegram,” the name is fine.
If this becomes AI monitoring for fast-moving information streams, I’d seriously pressure-test whether CoTel is too tied to the first channel.
I own Exirra.com, so I’m not pretending this is neutral advice. But I mentioned it because this is exactly the kind of product direction where a broader .com could become valuable before early demos, decks, and buyer memory start locking around the current name.
If it feels like a serious path, happy to compare CoTel vs Exirra privately from the buyer-positioning side, not just the name taste side.
Thanks — and if I decide to move in that direction, I’ll definitely let you know!
Makes sense. CoTel is clear for the current Telegram-first MVP.
If the professional/team monitoring direction becomes the main path later, that’s when the naming question becomes worth revisiting.
Good luck with the demo. That use case sounds stronger than the original personal workflow.