A Technical Advisor You Can Retain by the Hour — Like a Lawyer — Almost Doesn't Exist in the Market | Mizunara Co., Ltd. Case Study

Mr. Oda: Mizunara works on the theme of how to manage, edit, and publish large-scale documents more easily. In-house, we're researching and developing "Wordless" for large-document management and "OffiStill," an office tool for the AI era.

Real-world business documents don't work in plain Markdown alone. You need tables, figures, citations — a system that can handle them at a "production-grade granularity." This project was very much an extension of that theme.

Mr. Oda: We're a small team running both contract work and in-house product R&D in parallel. That's exactly why we emphasize not taking everything on ourselves, but bringing in strong external people for the critical pieces. Asking Leach to help was squarely in that spirit.

Mr. Oda: The biggest issue was that we had the requirement — "find similar passages" — but we couldn't decide which technology to choose within the team alone. The existing MySQL full-text search had clear expressiveness limits. Yet what should come next wasn't obvious.

When we first reached out, honestly, it was a pretty fuzzy request: "we'd like a way to find similar documents." Mr. Tominaga responded, "then vector search is a candidate." On our own, we hadn't reached that option.

Mr. Oda: If you only look at the search feature, there are plenty of choices. But in public-sector work, "if it's written in the spec, it must actually run, and the achievement rate cannot slip" is absolute. Even if a better idea comes up mid-project, doing something not in the spec breaks the Gantt chart, lowers the achievement rate, and the subsidy can be discontinued.

So we build the Gantt chart with plenty of buffer, prioritizing hitting ~100% on annual achievement rates above all. Within that constraint, we had to introduce a new element — similarity search. It came down to: could we pick not "the best tech" but "tech that fits the constraints"?

Mr. Oda: More than the implementation itself, we wanted a sounding board for design and architecture. Not enough demand to hire full-time. But if even a few dozen hours of a capable architect could come in, we felt the whole project would move forward in one step.

I knew Mr. Tominaga through our AIIT (Advanced Institute of Industrial Technology) connection — his reputation as an architect was strong, and his AWS expertise ran deep. He holds all 12 AWS certifications. At the time we consulted him, generative AI wasn't as developed as today; even searching the web for infra best practices had its limits. As someone who could fill that gap, he was essentially our only option.

Mr. Oda: What genuinely helped, once we started, was that he didn't just hand us a technical "answer" — he laid out the axes for how to choose, alongside the answer itself. His Slack responses were fast, so decision items never piled up.

The other big factor was fitting within the project's cost envelope. Even a ~¥60M project can't realistically hire a dedicated architect. That's exactly why "order design help only, for only the hours you need" was the right shape.

Mr. Oda: What I valued was that each option came with a cost estimate at our actual scale. There was a hard ceiling on server costs — cross it and we were done. With no dedicated infra person, being able to factor in operational weight as well was huge.

OpenSearch was an option, but we didn't need something that heavy this time. The value was choosing "the tech that fits these constraints" rather than "the strongest tech."

Mr. Oda: Honestly, the delivered PDF felt more than sufficient on its own. We had set an hour cap, but we didn't feel any need to burn through the remaining time. It felt more like, "I'd feel bad asking for more."

Mr. Oda: When we actually built it, validation and rollout went smoothly. What particularly helped was the pgVector install and initialization traps. "When and how to install in Docker vs. Aurora" is easy to get stuck on, but because we'd talked it through beforehand, we didn't.

Mr. Oda: For things like this, what helped wasn't "can we implement it" but "what's the natural way to wire it in" — and we had someone to discuss that with. Not just search; it felt like we'd gained a sounding board for the whole architecture.

Mr. Oda: Similarity search isn't a flashy feature, but it delivers cleanly on the need to find the right passage in thousands of pages at a natural granularity. Editors and reviewers search when they need to. It's a grounded use pattern.

Roughly a year in, operations are quiet — maintenance overhead is essentially nil. Document update frequency isn't high, so it's in the "quietly keeps running" phase.

Mr. Oda: Back when we consulted, generative AI wasn't as developed as today. Gathering best practices from the web was our best bet, and the part about judging them against our own constraints had to rely on people.

Now, generative AI has advanced and the number of technical options has, if anything, increased. But the more options you have, the harder it is to choose. Taking AI's answers at face value doesn't guarantee they'll fit the constraints of a live project. Knowing best practices isn't enough — you need someone who can pick up the constraints of a project that's already in motion and make a judgment call. That part still isn't easy for AI alone.

Mr. Kinoshita: Even when people are told "solve it with AI," many of them end up stuck right there. AI surfaces candidates, but to decide which to pick, you still need someone to talk it through with.

Deep domain knowledge and subtle constraints especially aren't well represented on the open web. The value of someone who can pick up the right context and translate it into decisions that fit the field doesn't fade as AI gets smarter — that's how I feel.

Mr. Oda: To begin with, someone you can retain as a technical advisor by the hour barely exists in the market. Consult only for the period you need, only on the topics you need — especially for small and mid-size development teams, it's a realistic option.

You get the knowledge of a highly capable expert at a cost lower than hiring. And with a ¥50,000/month starting price and short-term terms, the pricing is easy to wrap your head around. As a business model, it's genuinely solid.

Mr. Oda: This project isn't the end. We'd like to keep the relationship healthy and work together whenever a new question comes up. As generative AI continues to evolve and options multiply, a sounding board for design matters even more.

Mr. Kinoshita: The connection with Mr. Tominaga also runs through NAIST (Nara Institute of Science and Technology). Engineers from NAIST are genuinely excellent across the board, and the network keeps opening up new connections.

Mr. Oda: Once we were all out in the working world, I came to appreciate that caliber all over again. The quality of this engagement feels like a natural extension of that.

Mr. Oda: For anyone wrestling with technical challenges, just reaching out is a good first step. For people who understand the value of bouncing ideas off a capable architect, this service really lands. Lighter than hiring, but still able to engage deeply when it matters. That balance worked for us.