Conceived,
Built
& Running

AI orchestration that brings several frontier AI models together, cross-examines their reasoning and forges one answer.

• Four top models answer independently in parallel; a judge model checks evidence and contradictions, then integrates. Judgment, not majority vote
• 96.0% on GPQA Diamond (190 of 198), above the best published single-model scores
• 52.8% on Humanity’s Last Exam (400 questions), 1st of 33 models compared. Chat in JA/EN/ZH and an OpenAI-compatible API

* Both benchmarks are in-house measurements (June–July 2026). The HLE figure is not an official leaderboard score.

A voice AI assistant that runs your phone and surroundings just by talking. It holds a natural conversation and uses around 40 tools.

• It phones businesses for you, talks to them itself to book or check hours, and reports back — always after reading the plan back for approval
• Music, navigation, calendar, Telegram, posting to X, smart home, camera, shopping list and reports, all from one conversation
• Long-term memory you can search by voice, data kept separate per member, always-on Android app with over-the-air updates

A security assessment that attacks your site the way real attackers do, finds the holes, and explains the fixes in plain words.

• 40 checks in 9 areas; 20 of them run real but non-destructive attacks such as SQL injection, XSS and access-control bypass
• Attacks run only after site ownership is verified, cannot be aimed at anyone else’s site, and stop themselves if the server strains
• Reports with evidence and a fix for every finding, in JA/EN/ZH with PDF. A free one-minute check needs no sign-up

A research service that gathers reactions on X and has AI read them from five angles. Paste a post URL or keyword and get a report.

• Five modes: stance split, issue map, timeline, representative posts and a free-form question. It reads sarcasm and indirect criticism
• Up to 3,000 posts per run across replies, quotes and search, or 20,000 in a full scan, with a count-and-cost estimate before you start
• Before-and-after comparison around an event, without confusing correlation with cause. Results on screen and as a four-page PDF, in JA or EN

An end-to-end hiring system that re-measures people on six abilities that matter in the AI era, not degrees or titles. People make the final call.

• From a company URL and job title, AI researches the company, drafts the ideal-candidate profile and builds the interview
• Scores keep what candidates say apart from what can be verified; a high score backed only by talk is flagged “verify first”
• Only candidates worth checking are routed automatically to a live, AI-allowed task. Tested on 19 simulated candidates, including deliberate impostors

A system that seats 2 to 8 of 29 AI specialists for your topic, has them debate like a real meeting, and delivers a usable conclusion.

• Each panelist runs on a different vendor’s model (14 models from 8 vendors); agreement across vendors counts as stronger evidence
• A hidden director picks who speaks next and the chair checks convergence; consensus mode polls everyone and hands deadlocks back to you
• Conclusions come as conclusion, agreements, trade-offs, open issues and actions, exportable in six report formats

Tap once and talk. AI transcribes, tidies and sorts the memo, then files it in your own Google Sheet.

• Keeps recording with the screen off and re-sends automatically when the signal returns; nothing is lost even after a crash
• Learns your proper nouns to avoid mishearings and learns your own past labels to sort memos automatically
• Say “put it in Saitama” to Claude or ChatGPT and memos arrive from your AI chats too

A Fermi-estimation tool that gives an unknown number with its reasoning and range. AI builds the model; a maths engine does the arithmetic.

• Breaks a question into 4–12 independent factors with low/mid/high values and sources, then runs 10,000 simulations
• Shows which factors move the answer most, two ways; drag any value and the result recalculates instantly
• Reads Excel, PDF, PowerPoint and web pages as evidence, and writes a ready-to-paste explanation

A thinking partner that narrows a vague concern into the one question worth answering, built on the “start from the issue” method.

• Offers 3–5 candidate questions, each with a committed hypothesis, and tests whether each is truly worth answering
• Searches the web or runs a Fermi estimate mid-conversation, and records every decision on a live sheet
• A person always makes the final call, and who wrote or changed what is kept, so staff learn to think for themselves

A morning briefing that gathers only the news that matters to your company, says why it matters to you, and delivers it as five minutes of audio.

• Point it at your website and AI sets up your industry, competitors, partners and topics; it searches news in four languages
• Never repeats yesterday’s story; only real developments come back as updates, with the impact on Japan and on you
• Audio for the commute, email in Japanese and English, and Slack. Uses headlines only and always links to the source

A pronunciation clinic that fixes the English sounds Japanese speakers find hardest by measuring the voice itself, not by speech recognition.

• Measures vocal resonances (formants) and breath timing in the browser; thresholds were set on 444 labelled recordings
• Targets Japanese-specific habits such as R/L, TH and the extra vowel after final consonants, across 13 sound lessons
• 20 native-speaker videos with Japanese subtitles, and an AI coach that reads your measurements and history to pick the next fix

A training log that subtracts the heart-rate rise caused by heat and sun, so you see how much fitter you are really getting.

• Each Garmin session is joined with 10-minute temperature, humidity, wind and sunshine data from the nearest official weather station
• Removes heart-rate rise explained by heat, sun or a mask to give a corrected pace that compares sessions fairly
• Weight goals, AI analysis and coaching, and daily distance, step and workout bars drawn on your calendar automatically

A system that lets a racing yacht’s B&G H5000 instruments correct their own wind readings. The calibration table grows as you sail.

• Compares sailing before and after each tack or gybe; a correction is used only after passing fixed rules and an AI (Claude) review
• Writes are booked on shore and sent to the instrument only when connected and allowed, re-checked first and undoable in one step
• The same logs build sail crossover charts and polar tables, exportable to Expedition

An Android app that puts a yacht’s instrument data straight onto ENGO 3 smart glasses, so you see wind and speed without looking down.

• Built without the maker’s SDK: our own display driver, Bluetooth link and font
• Finds the instruments on the boat Wi-Fi automatically; a needle centred on the target angle shows any error at a glance
• Three screens for upwind, reaching and downwind, including values from Expedition. Connected on a real boat in October 2026

Research into Bitcoin trading strategies built around the roughly four-year halving cycle, plus the design of trading bots. Tested on 15 years of prices.

• Brute-forces thousands of settings and keeps only those that hold up when neighbouring values change, to avoid lucky picks
• The medium-term strategy closed all 20 trades in profit across the past four halving cycles (backtest)
• TradingView scripts, JA/EN reports and developer specs are ready. It does not trade today

* Backtest results on historical data. Not a promise of future returns.

A concept for a camera app: snap a few shots around you and AI reads the light and composition, then coaches you to a pro-looking photo on the spot.

• Spots usable elements such as light direction, puddles and branches, and suggests three ways to shoot
• Guides where to stand and how to hold the phone: “two steps left, phone at ground level, 3x zoom”