The story behind the search bar.
How a solo builder made a friend out of a recommendation engine.
March 2026
Origin
MILO started with a question that every serious reader has asked and no platform has answered well: why did I love that?
Not which genres it belongs to. Not how many stars it has. Not what other people with similar watch histories clicked on. Why it landed. What it stirred. The thing that made you stay up too late and think about it the next morning.
Every discovery tool on the internet knows what you’ve consumed. None of them know why it mattered. They see tags and metadata. You felt something. MILO was built to close that gap.
What MILO actually is
MILO is a manga discovery companion built by a single person — not a startup, not a team, not a company with a pitch deck and a Series A. One builder, working alone, because the problem was personal before it was a product.
The approach is different from what most platforms do. MILO doesn’t recommend based on popularity or collaborative filtering alone. It fuses multiple signals — content understanding, taste embeddings, community knowledge — and runs every decision through a question most recommendation systems never ask: would a real reader be happier?
There is no ad model. No engagement optimization. No dark patterns. No algorithmic tricks designed to keep you scrolling past what you actually wanted. MILO suggests. It never pushes.
How it’s built
MILO is built in public and with discipline.
Every new data source passes a provenance gate: is this source authoritative for this specific type of data? Not “does this field exist?” but “should we trust it?” Some sources are rejected. The reasoning is documented.
Every feature runs through controlled experiments with falsifiable hypotheses, statistical tests, and a question that sits above the metrics: does this actually make the experience better, or does the number just look good?
Every experiment — the ones that work and the ones that don’t — is recorded honestly. The results are never inflated.
This matters because the alternative is a recommendation engine that optimizes for clicks and calls it success. MILO is trying to do something harder: optimize for the feeling you get when a friend hands you exactly the right book.
The research
研究Building MILO generates real knowledge. Not proprietary secrets — genuine insights about how discovery works, how signals combine, where standard approaches fail, and what it takes to build something honest from scratch.
The builder behind MILO has chosen to share that knowledge openly.
This means publishing what’s learned along the way:
What happens when you analyze 2,300 community recommendation pairs and discover that niche titles get their only taste signal from explicit recommendations, while popular titles already have plenty from co-reading patterns. Where a new data source adds value isn’t always where you’d guess.
How to run controlled experiments as a solo engineer. Falsifiable hypotheses, paired statistical tests, per-stratum breakdowns that reveal where an approach works — not just whether the global number went up.
Every metric in recommendation systems is a proxy for something you can’t directly measure: whether the reader is happier. How to name your proxy, justify it, and stay honest about where it breaks down.
Not every data source that has a field is authoritative for that field. How to build a gate that catches this before bad data gets into your system.
What it actually looks like to build a product alone, with discipline, and without pretending the hard parts are easy.
These aren’t tutorials. They’re field notes from someone building a real product and paying attention to what actually works.
Values
Three words sit in MILO’s footer: Fast. Honest. Understandable. They’re not marketing. They’re engineering constraints.
The companion never makes you wait. Every query is cached, every fallback is graceful, every heavy computation runs off the main thread. Speed is respect for the reader’s time.
No ranking bias from affiliate commissions. No editorial voice that’s secretly sponsored. No content ratings quietly adjusted to increase reach. No engagement metrics shown to users to create anxiety. If MILO recommends something, it’s because the signals say you’d love it — not because someone paid for the placement.
The system behaves predictably. Search results don’t silently change quality. Settings do what they say. Errors explain what happened and what to do next. The reader is never left wondering why something looks different today.
MILO is being built slowly, carefully, and for the long run. Not to exit. Not to scale at all costs. To be genuinely useful to people who love reading and want to find something worth their time.
It’s just getting started.