We report from Tokyo, where Japan’s new disclosure rules are already producing results, and where the engine doing the work is our own.
A first-time customer proved the product. Underneath it sits something more durable: a platform powered by Straker’s own custom AI model.
The compliance clock is ticking. ASKA is ahead of it.
How one Japanese pharmaceutical company turned a regulatory mandate into a competitive edge.
From April 2025, the Tokyo Stock Exchange Prime Market required listed companies to disclose financial information in English simultaneously with Japanese: a starting gun for Japan’s 1,800+ Prime companies, every quarterly reporting cycle. Most are still finding their footing. ASKA Pharmaceutical Holdings moved first.
ASKA’s IR team completed its English third-quarter earnings report using SwiftBridge AI (Straker’s purpose-built multilingual service for investor relations) on their first attempt. It wasn’t just faster. It changed how the team works.
“Using SwiftBridge has made the English translation process 99% easier. The translations are of very high quality, and there is hardly any need for layout adjustments. I was able to spend my time on substantive revisions rather than on the translation work itself.”
“It’s as if we’ve gained an extra member of staff and everyone’s role has been elevated by one level. Staff who previously handled translation can now focus solely on checking.”
ASKA Pharmaceutical Holdings Co., Ltd. (February 2026). Translated from the original Japanese.With translation quality high enough that layout work all but disappears, time once spent on the mechanics now goes to substance, sharpening the disclosure itself. One team, elevated by one level. That’s the SwiftBridge effect, and it’s already selling.
This isn’t a faster workflow. It’s a better model.
The ASKA Pharmaceutical Holdings result is easy to read as a process win: automate the busywork, free up the team. It is that. But the reason the output needs hardly any human oversight and it clears a regulated IR disclosure on the first try because of the model quality, not a lighter review.
SwiftBridge AI runs on Straker’s own IR-specialised custom AI model, trained and tuned by us, not a thin layer over a general-purpose engine. That distinction is the business, and the numbers bear it out.
We benchmarked our model against the giants, and won. On unseen English–Japanese content (~6,000 held-out test points the model had never encountered), Straker’s custom model posted the highest overall quality score, ahead of a leading MT engine and multiple frontier models. “Unseen” is the point: this is genuine generalisation to new material, not memorised examples.
MT evaluation · en→ja · unseen benchmark
Straker takes the lead on Japanese translation
Scored equally across chrF, BLEU, BLEURT and TER on a held-out en→ja set. Bars carry 95% confidence intervals; higher is better.
Metric breakdown
| Metric | tiri-j-fin-7b-v1.2 | Global MT Engine | Frontier Model 1 | Frontier Model 2 |
|---|---|---|---|---|
| chrF | 62.40 | 50.84 | 47.91 | 40.08 |
| BLEU | 58.49 | 44.48 | 38.89 | 30.73 |
| BLEURT | 84.26 | 79.56 | 78.60 | 75.48 |
| TER | 70.45 | 59.54 | 50.70 | 48.06 |
| Weighted | 68.90 | 58.61 | 54.03 | 48.59 |
Source: Straker internal benchmark, June 2025. Evaluated on ~6,000 held-out English–Japanese segments not seen by any model during training. Weighted score is an equal-weighted combination of four standard machine-translation quality metrics (chrF, BLEU, BLEURT, TER); higher is better. Competitor systems were tested on their then-current publicly available versions as of May 2026.
Domain-specific by design. Generic models are fluent but generic. Straker’s model is tuned for the language that matters to our customers (Japanese-to-English financial disclosure, regulatory filings, IR terminology), so the first draft already speaks the register a listed company has to publish in. That’s why ASKA’s output needed “hardly any layout adjustment” on day one.
Built to verify, not just translate. Every disclosure passes a double-verification step (AI agents plus professional human translators), checking terminology, consistency and structure against the source. That’s the difference between “translated” and “publishable,” and it’s why the output stands up to regulatory scrutiny.
The model compounds. Every engagement sharpens our domain tuning. Competitors renting a third-party engine, the very engines our model just out-scored, can’t accumulate that advantage. We can.
Better model, faster compliance. Model quality isn’t an abstract score: it’s what makes the human side of the workflow efficient. When the model’s first draft is this accurate, reviewers verify and approve instead of rewriting, and the bottleneck of any regulated disclosure (human checking) shrinks. The result is concrete: SwiftBridge AI reduces the average time to produce a report by 70%, from ten days to three, with English executive summaries and timely disclosures delivered in as little as one business day: same compliance rigour, a fraction of the calendar.
This is what makes it a platform rather than a tool. SwiftBridge AI is the first application of the model stack, aimed at IR disclosure. The same proprietary models extend to regulatory, brand and specialist content wherever accuracy is non-negotiable, and the asset that powers all of it belongs to Straker.
SwiftBridge AI is the first proof point. The models are the moat.
At a glance
The customer: ASKA Pharmaceutical Holdings Co., Ltd. (first-time SwiftBridge AI user, English Q3 earnings disclosure)
The result: “99% easier”; high-quality output; minimal layout adjustment; team time redirected to substance
The trigger: TSE Prime simultaneous English-disclosure mandate, April 2025
The market: 1,800+ TSE Prime companies · quarterly disclosure cycle · active mandate
What powers it: Straker’s IR-specialised custom AI model + double-verification by AI agents and professional translators
Benchmark: #1 weighted quality score on unseen English–Japanese (68.90), ahead of a global MT engine (58.61) and frontier models (54.03, 48.59)
Workflow impact: Average translator report turnaround cut from 10 days to 3, as little as 1 business day for executive summaries
The model: AI for speed and volume; human reviewers for compliance review
The differentiator: We own and improve the model, not a workflow wrapped around someone else’s engine
The mandate is the proof of demand. ASKA is the proof it works. Our own model is what makes it ours to keep.
Last updated: 19 June 2026.