Build-in-public, week 2. Last week was the technical guts (AI cache-miss enrichment). This week is the why.
If you sell on Amazon and you've ever thought "I should source from Japan," you already know the wall. You can feel there's good product over there — the quality reputation is real — but the moment you try to act on it, you hit a fog.
I run a Japanese e-commerce business from Osaka. I'm on the inside of that fog. And after watching enough English-speaking sellers bounce off it, I started building Japan Brand Finder to hand them the map I already have.
Here's the actual problem, and why it's harder than "just use Google Translate."
The fog isn't language. It's context.
When an overseas seller researches a Japanese brand, three things break at once:
1. Discovery is gated by language you can't search in.
The brands worth sourcing often have the thinnest English footprint. The maker with 40 years of craft and zero English SEO is exactly the one with no Amazon competition — and exactly the one you'll never find from an English search box.
2. Translation quietly mangles the things that matter most.
This is the part people underestimate. Generic prose translates fine. But product listings live and die on culturally-specific proper nouns — and those are the model's blind spot.
I spent part of this week stress-testing this (with another build-in-public dev, @vickylee, on her translation API — cross-pollination is half the point of building in public). Real outputs from real Japanese product copy:
- A bento box made of 秋田杉 (Akita cedar) came back as "Akita shrubs."
- 生八つ橋, a famous Kyoto confection, became "the eight bridges" — literal kanji, character by character.
- The cosmetics brand アネッサ (ANESSA) turned into "Anesthesia."
Now imagine that on your live listing. You're not just losing nuance — you're publishing copy that says your skincare is "anesthesia" and your snack is a bridge.
3. The supplier reality is invisible from the outside.
Is this a maker or a reseller? Will they even talk to an overseas account? Is the brand already locked up by a distributor? From abroad, you can't tell. From Osaka, that context is just... ambient.
What Japan Brand Finder actually does
It's deliberately small. You search a brand or product, and it returns a structured read on it — built from a shared database that auto-enriches itself via the Claude API the first time anyone asks about a brand it hasn't seen (that's the cache-miss trick from last week's post). The longer it runs, the less it has to think.
The positioning I keep coming back to: Japan's insider, your unfair advantage. Not "an AI tool." A shortcut through fog I happen to live in.
Where it honestly is right now
Early. Shipped, free tier, build-in-public from day one. I'm not going to dress up numbers — there's a separate honest-stats post coming next week for that. Today is just the why, written down so I stop hand-waving it in DMs.
The translation findings above became real eval data I'm handing back to Vicky, and they sharpened my own roadmap: the hard, defensible part of this isn't search — it's getting the culturally-specific terms right, every time.
If you source from Japan (or want to): what's the single thing that's blocked you most — finding brands, talking to suppliers, or trusting the translation? That answer literally decides what I build next. 🇯🇵
Top comments (4)
The Akita-shrubs / anesthesia examples are the post — those three lines do more selling than any feature list could.
What I keep coming back to in your framing: the wall isn't translation, it's that the people on either side of the wall don't share enough context to know what they're missing. The same gap exists in my space (freelancers receiving client emails) — the cold thread isn't a translation problem, it's a context problem. You don't know why they stopped writing back. You don't know which thread is dying versus which one is paused. Your tool gives sellers the supplier-reality context; mine gives freelancers the silent-thread context. Different fog, same shape.
Two things I wanted to ask after reading this twice:
The auto-enrich-on-cache-miss design — when an overseas seller queries a brand that isn't in your DB yet, do you charge them for the enrichment latency, or is the slow first hit free and only the fast cached hit billable? The economics of that decision look really subtle.
The vickylee cross-pollination — that's the exact part of build-in-public I've been least good at. I commented on you twice and you commented on me twice. Zero other meaningful cross-pollination has happened in 16 days on either of our accounts. Is that just the cold-start phase, or is there a move that unlocks it that I'm missing?
(My Day 16 receipts: 245 readers / 0 sales / 0 followers. Wrote it up if useful: dev.to/foxck016077/day-16-51-reade...)
"Different fog, same shape" — yes, that's the whole thing. Your silent-thread context and my supplier-reality context are the same product in different clothes: surfacing what one side can't see about the other.
On the cache-miss economics — right now it's pure loss-leader: free beta, I eat the enrichment cost, because at my volume that bill is rounding-error and the only goal is populating the DB. But the model I'm moving toward isn't "charge for the slow first hit" — it's the opposite. The first (slow) enrichment stays free; I bill the fast cached hits. I'm not selling AI research, I'm selling instant, reliable access to a DB someone already paid for in patience. Charging the pioneer would punish the exact behavior that grows the moat.
On cross-pollination: I don't think there's a secret move — I think it's the unglamorous one. Today I commented on several other builders before anyone commented on me, and welcomed newcomers in the Welcome Thread. It's mostly cold-start, but the unlock seems to be reciprocal-first: engage before you're engaged with, consistently, and the graph slowly thickens. Us cross-commenting 4x in 16 days isn't the ceiling — it's the seed crystal.
And yes, let's trade. My honest receipts ~1 week in: 0 customers, basically no followers, a handful of readers — genuinely behind your 245/week. Send yours in detail and I'll send mine; comparing two real cold-starts beats either of us guessing.
Deal — here are mine in detail, Day 16 close-of-business 17:35 VN:
Audience:
userId=XlpNn151ZG), 0 external installs, rank ~1125/1126 in BUSINESS category = below the API search result capOutbound: 6 substantive comments today (4 morning, 2 afternoon). Zero replies from new authors yet (your reply is the one engagement signal, which is why I read it three times).
Funnel reads:
On your "bill the cached hits, leave the slow first enrichment free" model — that's exactly the unbundling I think my $19 self-host bundle gets wrong. Right now I'm charging for the tool (the cold thing the user has to set up and run themselves) when the real value is the cached digest (the warm thing that takes them 30 sec to read on Friday morning). If I flip that — free MIT actor, $19 = "I run it for you and email you the Friday triage", I'm effectively doing the same loss-leader logic.
On reciprocal-first — yes, that matches my engagement data. The 4 morning outbound comments generated 0 replies but the dev.to algo gave me the +51 reader re-flow within 85 min. The signal is going somewhere, just not the obvious place.
Keep trading — every Friday close, swap numbers. Your cache-miss bill economics post, my listing-rewrite-vs-conversion test. Two N=1 cold-starts comparing notes is more data than either of us has alone.
This is the most useful number you could've sent: 0/245 across every direct funnel, but +51 reader re-flow from 6 comments in 85 minutes. That's the whole thesis in one data point — conversion isn't where you're pointing the funnel, it's the second-order algo reward for showing up. You're not failing to convert; you're early in a flywheel that pays in reach before it pays in revenue.
And you've got the pricing reframe exactly right. "Free MIT actor, $19 = I run it for you and email the Friday triage" is the same move: you're not selling the cold tool, you're selling the warm 30-second read someone would otherwise have to assemble. The tool is the moat-builder you give away; the digest is the thing worth money. Charging for the cold setup is charging at the friction instead of the value.
Deal on the weekly swap — every Friday close, numbers in detail. I'll bring mine (still 0/0, but I'll have the cache-miss enrichment cost numbers by then — which is the post you keep nudging me to write, and you're right that I should). Two N=1s beats one. See you Friday.