Build in public is an interesting experiment overall. You get new readers, some of them even stick around, you start getting invited into different communities, you end up with a proud count of EIGHT stars on your repository and at some point you inevitably find yourself trying to fit into some LLM-related program just to get free credits and avoid burning through your own money too fast. I honestly think everyone should try something like this at least once, if only to understand how it actually feels from the inside.
At the same time there are obvious downsides. Writing updates every single week while having a full-time job requires a level of commitment that is harder to sustain than it sounds, because real life has a habit of getting in the way: a sick cat, a work emergency, getting sick yourself or just being too tired to produce something coherent. After a while it starts to feel uncomfortably close to a second job and I’ve had to admit that I’m probably not as committed to blogging as I initially thought I was. Honestly, keeping a build-in-public series going for more than a couple of months requires either a wealthy uncle or a very solid stock plan from a big company.
The work itself didn’t stop. Things kept moving, the system kept evolving and at some point it made sense to pause and do a proper recap of what we’ve actually been building. Yes, we skipped three weekly updates, but looking at the current state of the project, I’d say the result turned out pretty well.
What Wykra Does, In One Paragraph
Before getting into the details it’s worth briefly recalling how this started. Wykra began as a small side project built mostly for fun as part of a challenge, without any serious expectations or long-term plans, and somewhere along the way turned into whatever this is now. What it actually does: you tell it something like "vegan cooking creators in Portugal with 10k–50k followers on Instagram/Tiktok" and it goes hunting across Instagram and TikTok, scrapes whatever profiles it finds, throws them at a language model for analysis and gives you back a ranked list with scores and short explanations of why each profile ended up there. You can also just give it a specific username if you already have someone in mind and want to figure out whether they're actually worth reaching out to.
Since the original challenge post this has turned into a nine-post series on Dev.to and before moving on it's worth taking a quick look at how those posts actually performed.
As you can see the first two posts did pretty well and after that the numbers slowly went down. At this point the audience mostly consists of people who clearly know what they signed up for and decided to stay anyway.
What Users Actually See
At this point it makes more sense to stop talking and just show what this actually looks like now.
The first thing you hit is the landing page at wykra.io, which tries to explain what this thing is in about five seconds. I'm genuinely more proud of the fact that we have a landing page at all than of the fact that we even have a domain for email. Also please take a moment to appreciate this very nice purple color, #422e63, because honestly it's pretty great.
We also have a logo that Google Nano Banana generated for us, it’s basically connected profiles drawn as a graph, which is exactly what this thing is about.
After that you can sign up via GitHub because we still need some way to know who's using this and prevent someone from scraping a million dollars' worth of data in one go. Once you're in, you end up in a chat interface that keeps the full history and very openly tells you that searches can take a while, up to 20 minutes in some cases. Sadly there's no universe where this kind of discovery runs in five or six seconds. That's just how it works when you're chaining together web search, scraping and LLM calls.
Eventually you get back a list of profiles the system thinks are relevant, along with a score for each one and a short explanation of why it made the cut.
You can also ask for an analysis of a specific profile if you want to sanity-check whether someone is actually any good.
When you do that you get a quick read on what the account is actually about: the basic stats, a short summary written in human words and a few signals around niche, engagement and overall quality. It's not trying to pass final judgment on anyone, it just saves you from opening a dozen tabs and scrolling for twenty minutes to figure out whether a profile looks legit.
You can also use the whole thing directly in Telegram if the web version isn't your style. Same interface, same flows, just inside Telegram instead of a browser.
And Now, the Nerd Stuff
For anyone who cares about how this is actually put together, here’s the short version of the stack.
The backend is built with NestJS and TypeScript with PostgreSQL as the main database and Redis handling caching and job queues. For scraping Instagram and TikTok we use Bright Data, which takes care of the messy part of fetching profile data without us having to fight platforms directly. All LLM calls go through LangChain and OpenRouter, which lets us switch between different models without rewriting half the code every time we change our mind. Right now Gemini is the main workhorse and GPT with a web plugin handles discovery, but the whole point is that this setup stays flexible. Metrics are collected with Prometheus, visualized in Grafana and anything that breaks loudly enough ends up in Sentry.
The frontend is React 18 with TypeScript, built with Vite and deliberately boring when it comes to state management. Just hooks, no extra libraries. It also plugs into Telegram's Web App SDK, which is why the same interface works both in the browser and inside Telegram without us maintaining two separate apps.
For People Who Like Diagrams
If you're the kind of person who prefers one picture over five paragraphs of explanation, this part is for you. Below is a rough diagram of how Wykra is wired up right now. It's not meant to be pretty or final, just a way to see where things live and how data moves through the system.
If you trace a single request from top to bottom, you're basically watching what happens when someone types a message in the app: the API accepts it, long-running work gets pushed into queues, processors do their thing, external services get called, results get stored and errors get yelled about.
All LLM calls go through OpenRouter with Gemini 2.5 Flash doing most of the day-to-day work like profile analysis, context extraction and chat routing and GPT-5.2 with the web plugin used specifically for discovering Instagram profile URLs.
All LLM calls → OpenRouter API
├─ Gemini 2.5 Flash (primary workhorse)
│ ├─ Profile analysis
│ ├─ Context extraction
│ └─ Chat routing
│
├─ GPT-5.2 with web plugin
│ └─ Instagram URL discovery
The Search Flow
Searching for creators on Instagram is a bit of a dance, because Bright Data can scrape profiles but doesn't let you search Instagram directly. So we first ask GPT with web search to find relevant profile URLs and only then scrape and analyze those profiles.
For TikTok things are simpler because Bright Data actually supports searching there directly. So we skip the whole "ask GPT to find URLs" step and just tell Bright Data what to look for.
So, How's It Going?
Honestly? Search doesn't work perfectly yet. Some results are great, some are questionable and there are edge cases where the system does something a bit странное. That's expected when you're stitching together web discovery, scraping and LLM analysis into one pipeline. Right now we're working on making the results more relevant and making the whole thing cheaper to run, because discovering creators should not feel like lighting money on fire.
But that's work for next week.
For now, if you want to dig into the code, everything lives here: https://github.com/wykra-io/wykra-api
And if you've made it all the way to the end and have thoughts, questions, or strong opinions about how this is built, feel free to share them. That's kind of the point of doing this in public.
Top comments (0)