Qualiteg at Startup JAPAN EXPO 2026: Exhibition Report
Hello!
This is the Qualiteg Business Development & Marketing team!
For two days, from Wednesday, April 15 to Thursday, April 16, 2026, "Startup JAPAN EXPO 2026" (hosted by Eight / Sansan, Inc.) was held at Makuhari Messe, and
Qualiteg, Inc. was there as an exhibitor!
Startup JAPAN EXPO 2026: Exhibition Overview
| Item | Details |
|---|---|
| Dates | Wednesday, April 15 – Thursday, April 16, 2026, 10:00–17:00 each day |
| Venue | Makuhari Messe, Exhibition Halls 7 & 8 |
| Booth | 16-16 |
| Organizer | Eight (Sansan, Inc.) |
Our theme for this exhibition —
"Delegate in parallel. Never stop thinking."
Our flagship product "Bestllam®" recently gained a new AI Agent feature, and along with our AI security solution "LLM-Audit™", we ran live demos of both throughout the two days.
Our booth number was 16-16. We set up camp in a corner of Exhibition Halls 7 & 8, fired up as always 💪
Press Release
株式会社Qualiteg
The Day Before [Day 0] — The Eve: "We Will Not Be Late Again" — the Forward Base Is Born🏨
Please let me start this report right here.
Some of you may remember our previous CEATEC 2025.
Yes — "arrive last, stay until the very end" — the legendary edition in which we accidentally pioneered that whole new genre...
Late-night Makuhari Messe, an empty exhibition hall, Orion shining overhead 🌌 — a beautiful sight, and one we never want to see again 😇
So this time, we made a resolution.
"This time, we stamp out lateness and oversleeping (💦)"
In the name of that resolution, we arrived a day early.
At the APA Hotel & Resort Tokyo Bay Makuhari — right next to Makuhari Messe and within walking distance — we decided to set up a "forward base" for the event.
From the forward base, Makuhari Messe is in perfect view below (and that is ZOZO Marine Stadium 🥎 in the upper left, by the way — a game was underway ^^)
On site, we ran a squad of three energetic marketing team members, with engineers along as well — a lineup ready to handle any technical issue on the spot.
"Surely this is bulletproof."
...or so we thought at the time.
The demos, as it turns out, had "stage machinery" behind them
That night after arriving, in a room at the forward base, we were once again rehearsing the full two days of demos.
Let me take a brief detour into what goes on behind the demos.
What we absolutely wanted the Bestllam® AI Agent demo to convey was —
"the sense of an agent genuinely woven into the daily operations of a Japanese company"
.
"An AI agent that searches the web and writes summaries" — these days, that alone is not exactly rare.
What we wanted to show goes one level deeper:
"an agent that connects to core business systems and sales databases and works inside real workflows."
So for this demo, our engineering team had built
a set of "mock MCP servers" mimicking corporate core systems and sales-management databases
from scratch 🛠️
- A fictional customer master (thousands of records)
- Fictional sales transactions
- Fictional contract data and confidential project names
- Fictional attendance data
- Fictional business documents: contracts, meeting minutes, customer-support notes, and more
All of it was designed to answer with the behavior of a real production back-end — stage machinery we planned to run quietly behind the booth.
With this in place,
- in the Day 1 Bestllam demo, the agent could take an instruction like "aggregate last month's top five sales by department, line them up against next month's forecast, and put it all into PowerPoint," query the mock databases, and return results
- in the Day 2 LLM-Audit demo, we could take "obviously confidential"-looking text pulled from the mock core system and, the moment it is about to be sent to an LLM, detect, mask, and pseudonymize it in real time
— which is how a single narrative spanning the two days holds together.
And then, as night fell, the mock MCP got moody
The engineer rehearsing at the forward base muttered:
"...Wait. Is this behaving inconsistently?"
The symptoms were the quiet, nasty kind.
- When multiple sub-agents accessed it in parallel, the response order would occasionally scramble
- For certain query patterns, the returned data schema would subtly drift
- As a result, the agent would insert a "just-in-case retry," adding a few seconds of perceptible lag
The products themselves (Bestllam / LLM-Audit) were working flawlessly. We had verified them again and again in the studio environment at HQ.
The moody one was the mock core-system MCP we had built ourselves for this demo.
But from a visitor's point of view, none of that backstory matters.
"The agent seems to be... waiting on something."
— the instant it looks that way, the persuasive power of "Delegate in parallel, never stop thinking" erodes a few tenths of a second at a time.
A trade-show demo forgives not even a fraction of a second of awkwardness.
"Well then. Let's get to work." — sekihan onigiri, fuel for the soul 🍙
An emergency war room sprang to life in a room at the forward base (APA Hotel Makuhari) 🔧
- Reworking the mock MCP's parallel-processing locks
- Pinning down the response schemas
- Mapping out the likely instruction patterns and pre-rehearsing the responses to each
...Before we knew it, we were starving.
The Lawson convenience store on the hotel's 2nd floor — we made a dash for it and bought a huge pile of sekihan (red bean rice) onigiri 🍙
"Why sekihan?" If you ask, there is no real reason. The moment it caught our eye, the "good-luck ritual" switch flipped on.
"Tomorrow is our first battle, after all."
On that mysterious justification, deep into the night at the forward base, we kept working while happily devouring sekihan onigiri.
(All sekihan calories were, for accounting purposes, booked as "fuel for the soul." Nutritionally speaking, no comment 😇)
Munching onigiri, hitting the mock MCP, reading logs, fixing again.
Unless the "stage machinery" runs perfectly, the true appeal of the stars (Bestllam / LLM-Audit) never comes across — with that in mind, every member (including the engineers joining remotely) kept their hands moving.
Before we knew it, it was dawn.
We happened to pull back the curtain, and —
Beyond ZOZO Marine Stadium, across a Tokyo Bay just beginning to pale, something like a tower floated faintly in the distance —
it was Tokyo Skytree!
It looked almost divine!
The slightly damp morning air.
The hushed stillness peculiar to daybreak spread out below us — it looked like the stage where the coming battle would be fought.
When the mock MCP finally settled into stable operation, the engineer allowed himself a satisfied word:
"All right. The stage is set. I think we can pull off day one."
Armed with that mysterious confidence, we headed for the Day 1 battlefield.
[Day 1] Bestllam AI Agent — the moment your "dedicated team" springs into action
The morning's cloud cover broke into pleasant sunshine right around the time the booth opened 🌞
Just like the weather, our Day 1 booth got off to a smooth, energetic start.
The Day 1 headliner: Bestllam®'s new AI Agent feature.
The conventional AI chat experience is bound to a turn-based structure: the human waits for the AI's reply before thinking about the next step. The more you use AI, the more your thinking gets chopped up — you know the feeling.
Bestllam's new AI Agent turns this upside down.
The moment you give an instruction, your own dedicated team springs into action.
A supervising agent acts as the control tower while multiple sub-agents work in parallel, asynchronously handling research, deck building, data cleanup, email drafts, and more. You can keep stacking new instructions without waiting for replies.
- "Aggregate last month's top five sales by department and line them up against next month's forecast"
- "In parallel, draft the visit schedule for our key accounts"
- "While you're at it, condense those visit plans into a one-page report for the department head"
...and so on — you can already keep piling on instructions like this.
Reactions to the AI agent were especially enthusiastic among executives and business leaders, and when we heard
"We use AI internally, but it still felt like working alone. This — this is a team."
— the exact dead-center comment we had been aiming for — we struck a victory pose on behalf of the development team 👊
"An AI agent that works plugged into your core systems" — that was the moment we felt the message truly land.
Thanks to everyone, our Day 1 booth was wrapped in energy and excitement.
Members who were a bit worn from days of preparation drew energy from the visitors' reactions, and we kept the demos running until the closing announcement echoed through the hall.
Find Qualiteg!
The whole venue, for that matter, was buzzing too.
Here is what the hall looked like on Day 1. Impressive, right?
So, a quiz.
Where is the Qualiteg, Inc. booth?
Hint: wherever the energy for deepening creativity is being unleashed the hardest.
Installing an AI chat tool alone gets you nowhere near true operational improvement
Allow me to also touch on the consulting services we offer.
At this event too, visitors who stopped by or booked meetings shared challenges such as:
"We introduced generative AI, but chat alone is not translating into efficiency gains"
"We get as far as a PoC, but company-wide rollout is another story"
as well as:
"We want to overhaul our information systems development with generative AI, but we cannot see the whole picture — where to start, or what changes our development process needs"
and other voices along those lines.
This is exactly our home turf, but to be candid:
Simply building a tool or SaaS, handing it over, and saying "here you go" rarely produces real transformation.
At Qualiteg, consultants with years of experience at global consulting firms team up with our AI specialists to support AI-driven operational improvement that leads to genuine change.
Visitors with precisely these concerns came by, and we promptly received requests for the free spot-consulting sessions offered to event attendees. To mark the exhibition, we have expanded the number of free spot-consulting slots — if you are interested, please contact us here.
The answer: Qualiteg is right here
So, how was our little Find Qualiteg challenge?
Was it easy?
Here is the answer.
To everyone who found our booth among so many and stopped by or booked a meeting — thank you once again.
[Day 1, Night] The "confidential-document endpoint" acts up again 🌙
Riding the elevator back to the forward base after Day 1, the booth crew was brimming with a sense of accomplishment.
"Tomorrow, we go again!"
That's right — the star of Day 2 would be LLM-Audit™.
It works upstream of confidential information ever reaching the LLM, performing real-time detection, blocking, masking, and pseudonymization. For example,
"We received an inquiry from Taro Tanaka (090-1234-5678)"
— the moment someone tries to send that to an LLM, it turns into
"We received an inquiry from [Customer A] ([phone number])"
— transformed in an instant (a few milliseconds). That "pleasant feeling of never noticing the security check is even there" — being properly secure while keeping security invisible during AI use — is the very soul of this product's UX.
And to make this demo look genuinely real, we needed
(no real confidential data allowed, naturally) business documents that look every bit like confidential material —
pulled straight from the mock core-system MCP.
The lightweight endpoints were fine — but the "long documents" were untested
The mock MCP we fixed on the night of Day 0 covered the lightweight-data endpoints — sales aggregates, schedules, customer IDs... that structured data got us through Day 1 somehow.
However, contracts, minutes, customer-support notes, and internal approval documents — the endpoints returning such long-form business documents had not been called even once during Day 1.
Rehearsing for Day 2, we hit that confidential-document endpoint — and then:
"...Ah."
The engineer's face clouded over again.
The symptoms were, once more, subtle — and this time fatal.
- Sometimes the long text contained no PII at all (leaving the demo with no showpiece moment)
- As a result, the LLM-Audit detection engine would trigger at times — and at other moments things would occasionally slip right through
- We simply did not have nearly enough variety in the prepared confidential documents — no matter how many times visitors tried it, the same minutes and the same contract kept coming up
LLM-Audit's own detection engine (PII-Fi™) was working flawlessly.
The moody one, yet again, was the mock core-system MCP. And this time it had stumbled at the most critical part of all — "fetching the confidential data".
"Again?! We just did this yesterday...!!"
We looked up at the heavens 🙄
But there was no way we could walk into Day 2 with this unresolved.
The LLM-Audit demo lives on that jolt of "wait — it found it even in prose that natural!". Text with no PII at all gives it nothing to find; yet randomly generated, obviously mechanical text makes detection feel so inevitable that it is simply boring.
Natural Japanese business documents, the kind that actually circulate in real work — we needed a fairly large stock of exactly that.
"Well then. Let's get to work." — night two in a row
The forward base war room was hereby approved for a second consecutive night of operations.
Last night we fixed the mock MCP's skeleton; tonight we would finish the content drawn out of it.
Last night was sekihan. Tonight —
"Pizza. Let's do pizza."
We dashed to the late-night restaurant inside the APA Hotel (MOMO CAFE) and got pizza to go 🍕
"This pizza is way too good..."
Alongside last night's sekihan onigiri, it took its place as late-night "fuel for the soul". As for the calories in midnight pizza — see the earlier note.
Between slices of pizza, the work looked like this:
- Contracts, minutes, customer-support notes, internal approval documents — writing out fictional variations, one after another (authored by the marketing team). Sure, we could have generated them with an LLM through Bestllam, but where would the fun be in that?
- Salting the documents with details like 「田中太郎様」 (Mr. Taro Tanaka), 「株式会社〇〇商事」 (○○ Trading Co., Ltd.), "090-1234-5678", and 「第3四半期決裁案件 — 社外秘」 (Q3 approval item — internal only) — realistic embedded data mixed in naturally
- We also deliberately worked in spelling and notation variation — the engine is robust to that. And since the marketing staff do not know the internals of the PII logic, the text conveniently avoids being easy to detect. All the better.
- Finally, on the LLM-Audit side, tuning the perceived speed of detection, pseudonymization, and display to look one notch more satisfying
Funny thing: preparing for a trade-show demo like this often pushes development further than ordinary office days do. Having seriously exhibited at more than ten shows, we have learned that even when a hands-on demo does not quite come together, imagining the real thing this concretely and demoing it is genuinely important.
4 a.m. — "Hang on — isn't this URL wrong?"
It was just past 4 in the morning.
A room in the forward base. Our wrestling match with the mock MCP and the fictional minutes was finally nearing the finish line.
The engineer writing code idly picked up one of the flyers meant for booth handouts from the edge of the desk.
"...Hang on. Isn't this URL wrong?"
The air in the room snapped frozen 🧊
"Huh." "Huh?" "Whaaaaat!?"
Everyone rushed to the desk.
We checked, and —the QR code and destination URL introducing LLM-Audit™ — the Day 2 feature — had an error.
Day 1 had centered on Bestllam, so nobody had noticed. But the star of Day 2 was LLM-Audit.Hand these out as-is, and we would be sending visitors to the wrong page.
At a trade show, flyers and posters are your face — your business card. Worse, this was the very destination for people who took an interest in LLM-Audit™ — in other words, the doorway to the information we most wanted to deliver was wrong.
"We cannot stand at the booth tomorrow without fixing this."
The verdict was unanimous.
"Accea, we're counting on you...!!"
We corrected the poster and flyer data at full speed. Right URL, right QR code, right destination. At 4 a.m. in the forward base, one more director's check, a proofread, a final review.
And our lifeline was the Accea Kaihin-Makuhari print shop, near Makuhari Messe.
Opening time: 8 a.m..
"If we submit online now, and they notice it right at 8:00 when they open, it could be printed by 9... maybe." "Doors open at 10... that is razor thin." "Let us pray."
We loaded the data into the submission form and, fingers resting on the send button, offered up a genuinely earnest prayer together 🙏
"Accea, we're counting on you...!!"
After the all-nighter — and then, clear skies
Only after all that did the forward base war room finally close up shop.
- Mock MCP confidential-document endpoint → fixed ✅
- Fictional business-document variations → written ✅
- Poster and flyer URL fix → submitted for printing ✅
Depositing dark circles under our eyes, we nevertheless felt a strange sense of fulfillment.
That fulfillment surely comes from doing this because we want to — not because we have to.
And then we looked up, and —
morning, again.
We opened the curtains, and —
Well, look at that — Mt. Fuji 🗻 was visible.
A complete reversal from the previous day's overcast: a faint Mt. Fuji across Tokyo Bay,
and the roof of Makuhari Messe glittering in the reflected light
"All right. Day 2. We can do this."
Two consecutive nights of emergency work behind us — and yet the cloudless sky somehow felt like a hand at our backs ♪
[Day 2] The LLM-Audit demo — the heartfelt truths of AI security
Day 2 belonged to LLM-Audit™.
But first — the first-thing-in-the-morning call from Accea 📞
Walking toward Makuhari Messe under a clear morning sky, one of our members felt a phone buzz.
"We rushed the data you sent — it's already printed. Come pick it up anytime!"
...We stopped in our tracks for a moment.
"You... really did it...!?"
The files we had submitted with a prayer in the dead of night had been printed the moment the shop opened — already finished.
You could tell that the person on the other end of the line had completely grasped our "we need this by the skin of our teeth" situation and thrown themselves into the job.
The members commuting by train picked up the posters at the Accea shop near the station, and disaster was averted.
To the staff of Accea Kaihin-Makuhari — thank you, truly, from the bottom of our hearts 🙇♂️🙇♀️
That sudden overnight submission, with barely any explanation attached — and still they read our intent, finished it first thing in the morning, and called first thing in the morning. That is the spirit of a true professional, and it moved us.
Clutching the freshly corrected flyers and posters, we slid into our Day 2 booth.We made it.
And now: LLM-Audit takes the stage
Freshly printed, correct flyers and posters. Our preparations continued, quite literally, until the very last moment 🙏 (To our past selves of the last two nights, and to Accea — thank you, truly.)
With LLM-Audit (PII-Fi):
- "Text containing phone numbers or addresses" → masked in an instant
- "Customer name + contract value" → pseudonymized and passed to the LLM with context intact
- "Internal-only keywords" → instantly detected and blocked by the pre-trained PII-Fi™
- "Names with honorifics and project names blended naturally into meeting minutes" → detected, including Japanese-specific orthographic variation
— all executed in a flash (a few milliseconds).
And then the "quietly heartfelt worries" started coming out
What left the deepest impression this time were the "quietly heartfelt worries" we heard from the people who visited.
"Our employees are already pasting meeting minutes into ChatGPT on their own..." "We found out contracts were being summarized with AI, and our legal team turned pale" "Shadow AI — how are we even supposed to stop it?"
...Allow me to be serious for a moment here.
When we first started building LLM-Audit, the use case we had in mind was defending LLMs against malicious attackers.
That is: attacks on your own AI (jailbreaks), inappropriate outputs, hallucination detection — that family of problems.
(Video) The problem LLM-Audit originally set out to solve
But as development went on,
we became convinced that the truly urgent problem is
data leakage from well-intentioned, everyday work
— the one that cannot wait.
(Of course, LLM-Audit also holds its own against jailbreaks and prompt injection, and our AI red team sharpens its skills day and night 💪)
This time, we heard exactly those real voices from the field, and plenty of them.
- Accidentally sending confidential information to an AI
- Employees adopting personally contracted AI tools because they are convenient (shadow AI)
- By the time anyone notices, there is no way to trace where the information went
These are precisely the territory LLM-Audit is meant to cover.
"So our direction was right after all."
There was a moment in the booth when the team quietly exchanged glances and nodded 🥹
It exists to support exactly that "final step" — a truth that sank in all over again.
[Day 2, Evening] That sunset from the forward base 🌇
And so,
our long two days of exhibiting reached closing time.
We packed up the booth, gathered our things, and made it back to the forward base by evening.
Riding the elevator up to the high floors, we happened to glance out the window —
It was breathtaking.
The sun sinking beyond Tokyo Bay.
The moment orange melts into marine blue.
"We really ran the full two days, didn't we."
"That pizza was so good."
"Accea really saved us."
"That one visitor — they laughed so much."
We watched the sunset in silence for a while.
A trade show — from preparation through showtime to teardown — really is a "drama."
The Day 0 night of sekihan onigiri, the Day 1 clouds, the buzz of the crowd, the midnight pizza on Day 1, the endless writing of "fictional minutes," the late-night poster incident, the first-thing-morning call from Accea, the cloudless morning sky, the quietly heartfelt words from customers, the energy of the Horiemon keynote — all of it, together, made up these two days.
Maybe this sunset was a reward shown only to those who run all the way through 🥹
Wrapping Up — A Feel for What Comes Next
This Startup JAPAN EXPO 2026 lasted only two days, but for Qualiteg they became two days of real significance.
A bigger response to the Bestllam® AI Agent than we imagined
The concept of "Delegate in parallel, never stop thinking" landed deeply, as something visitors could feel firsthand — that was our clear takeaway.
The passionate reactions from executives and business leaders in particular gave us conviction that the "AI that works" direction we are pursuing is exactly what Japanese business floors are asking for.
Real-world expectations for LLM-Audit™
Shadow AI and accidental data leakage — these are no longer "future concerns" but problems unfolding right now, in front of us. The quietly heartfelt voices we heard from visitors showed that AI adoption in Japanese companies is moving into its next phase — governance and speed, achieved together.
LLM-Audit™ is the product that serves as that "last line of defense." Hearing "Yes — this is exactly what we wanted" during the demos was the greatest encouragement we could ask for.
And the feeling of a team that ran the whole way
This Startup JAPAN EXPO was an all-hands effort by the marketing and engineering teams.
Behind the smiles at the booth were two nights of behind-the-scenes work, midnight fixes, dawn final checks, and countless preparations.
Even so, we finally broke the "arrive last, stay until the very end" jinx this time! ✨
Next time, we will surely reach the perfectly ordinary genre of "arrive first, leave first" as well (can we, though?).
To everyone involved, and everyone who visited
To the organizers at Eight (Sansan, Inc.), the venue operations staff, the team running the concurrent Climbers event, and every single visitor who stopped by booth 16-16 —
thank you. Truly, deeply, thank you 🙇♂️🙇♀️
And to the staff of Accea Kaihin-Makuhari —
a sudden pre-opening submission, an unreasonable schedule — and still you printed it first thing and called first thing. That professional response has our deepest gratitude. Thanks to you, we made it in time for the Day 2 opening.
Incidentally, the Day 1 posters were also a rush job handled by the Accea Otemachi branch — without Accea, this exhibition would never have come together🙏
"We want to try the Bestllam agent in our department" "I want to share the LLM-Audit story with our IT department" "Let us start with one consulting session with Qualiteg"
— we received so many comments like these. We will follow up on every one of these connections, one by one, and carry them forward.
And to everyone who traveled from afar to find our booth, everyone who cheered us on over social media, and you, who read this report to the end:
Under our guiding principle of "deepening the creativity of people around the world", Qualiteg will keep building AI that genuinely blends into the Japanese workplace.
The whole team looks forward to meeting you in person at our next event!
Thank you very much for reading to the end ✨
See you next time! 😊
