If you skip this
A folder dump with no tags. Search leaks or returns text nobody can source.
RedRobot Tokyo Worldwide
Production software for the AI decade, built to run on your systems.
RedRobot is a boutique studio that consults and builds production AI, data platforms, and enterprise blockchain. Headquarters in Tokyo. Clients worldwide. We write down what success means, ship a first real version on your infrastructure, measure it, harden it, and hand it over so your operators own the next mile.
AI · Data · Blockchain · Security / keys · Worldwide
How engagements usually feel
-
First real version, written clock
Thin scope on your systems. Success criteria and timeline agreed up front.
-
Built for scrutiny
Sources, keys, residency, and approvals you can show in review.
-
Your team runs it
Handoff is the product, not a perpetual secondment.
-
Tokyo base, global work
Roppongi studio. Delivery for clients worldwide.
RedRobot
Experience is the advantage AI cannot fake.
The market is racing hard on AI right now. New models arrive weekly. Judgment does not. Our team has spent thirty-plus years on the floor of each major wave: personal computing, the commercial internet, mobile, open-source data at scale, crypto-era infrastructure, and now production AI. We have kept networks up, run systems of record, sat in research labs (including France Telecom R&D and NTT DOCOMO R&D), and shipped software that had to work outside the demo room. That is what we bring when we design agents, pipelines, or chain systems for your environment: what fails under load, what auditors ask for, and what your team can still run after we leave.
Entity
RedRobot K.K.
Based in
Tokyo
Focus
AI · Data · Blockchain
What we build
Three planes. One production habit.
-
01
Data
Systems of record, pipelines, access, and lineage, so AI and operations sit on truth you can recover. Schema, identity, backup, and failure domains come before any model conversation.
-
02
AI
Answers and agents on your live data: retrieve evidence first, keep permissions at the data, use the smallest model that works, require a human when the step cannot be undone, and leave a log you can open.
-
03
Blockchain
Enterprise L1/L2 when operations need a chain that talks to real systems, not market theater. Code and contract review, observable settlement, and key or post-quantum paths when the threat model demands them.
How production AI fits together
Open building blocks. Your data. A path you can actually run.
Most production AI pain is not the model. It is the mess around it: files with no clear source, search that ignores who may see what, answers you cannot prove, and risky steps with nobody on the hook.
We build that path with open tools people already trust (MySQL, Postgres, NATS, OpenSearch, and peers), fitted to your cloud and rules. Click any box in the diagram for what that stage does.
Click a stage
06 stages
01
stage
Get the material ready
Manuals, tickets, and feeds are cleaned and tagged with source and access rules before anything is embedded or generated. Rights start when data is written.
Cost note
Prep once. Re-index when content changes, not every user turn.
Think of it as a factory floor for answers: raw material in, durable machinery, a controlled store, a careful readout, then a human seal when the stakes are high.
-
Tools you can hire for
MySQL, Postgres, and other public open-source pieces your team can hire and run.
-
Built on your estate
Wired to your login, data, and cloud. Not a demo sandbox you throw away.
-
Spend you can forecast
Private models by default. Large public models only when policy and need agree.
-
You own the keys
After handoff, your operators run it. We return when the next scope is worth it.
Next
How we engage, and path A vs path B cost.
MySQL-class open tools · Your permissions · Proof you can open · You run it after handoff
How the team delivers
Same control loop on every plane.
AI on real data, data platforms under load, or chain that touches money and keys. We use one loop: write success criteria, ship a first real version on your estate, measure, harden, hand over. We expand the design that already clears the bar rather than discarding a pilot and starting over.
-
01
Define success
Write accuracy, cost, residency, latency, and audit needs. For chain: who signs, where keys live, and what finality means. If it is not written, it is not a requirement.
-
02
Ship on your systems
Live identity, data stores, and tools, not a disposable sandbox. Scope is thin; APIs, schemas, and ACLs are real. You still set product direction.
-
03
Measure, harden, hand over
Groundedness, cost per successful job, and error budgets (or chain finality and custody SLOs). Then monitoring, human gates on irreversible steps, runbooks, and a clean handoff. Your operators run it.
How a project usually moves
One method across AI, data, and chain.
Whether we wire AI into your databases, repair a data plane, or stand up enterprise chain infrastructure, the discipline is the same: define success, ship a first real version on your systems, measure against written criteria, harden, hand over.
-
01
Define success
Accuracy, latency, cost, residency, and audit, written before tools. For chain: signers, key custody, finality, and failure domains.
-
02
Ship a first real version
On your identity, data plane, and network. Scope is thin. Contracts (APIs, schemas, ACLs) are real. Models and frameworks are tools. We own the design.
-
03
Measure
Groundedness, recall, cost per successful job, error budgets. For chain: finality and custody SLOs. Kill or promote with evidence, not hope.
-
04
Harden
Observability, human gates on irreversible actions, runbooks, load and failure drills. Expand only the design that already clears the bar.
-
05
Hand over
Credentials, docs, and an operational walkthrough. Your team runs it. We return when the next scope is worth a new loop.
Selected work
What changed after it ran in production
Client names stay private. The problems do not: cost that would not sit still, data that left the building, or records nobody could audit. Here is what the floor looked like before and after.
- 01
Field service · APAC
Private SOP assistance: about 70% less cost per ticket than a public wrapper
14 weeks to handoff; ops runs it now
- ~70% lower cost per ticket vs public GenAI wrapper
- Customer and field data stay inside your perimeter
- Token use stays predictable under load
What was failing
Technicians needed the right procedure from manuals and past tickets while they were still on site. The default product was a thin app over a public model: paste SOPs and history, hope for a clean answer. As volume grew, bills jumped without warning. Worse, procedures and customer context left an environment you control. Answers drifted from the official SOP. Compliance could not reconstruct a single call from evidence.
What the team built
We kept manuals and tickets in a store the client owns. Retrieval pulls only the pages that matter; generation runs on a private path inside their environment. Smaller models handle routine load; low confidence escalates to a person. Every material step is logged. Token design follows evidence first, not stuffing an entire manual into every prompt.
What operators got
Faster answers that cite the correct SOP revision. Cost per ticket fell about seventy percent versus the public wrapper baseline. No surprise spend spikes when traffic rose. Field knowledge and customer data stayed internal. Ops and compliance can open the trail. The stack holds under load, and they own it after handoff.
Where AI helped
Retrieve and draft on private data; escalate hard cases. Does not invent procedures or export those documents to a public model path.
Where software carried it
Jobs, storage, permissions, packaging, gates, logs, and runbooks. An internal system your team can operate.
- 02
Documentary / non-fiction post · video
Find the usable interview moment without days of scrubbing rushes
12 weeks; post owns the library
- Multi-day scrub sessions → minutes of guided search
- Every candidate carries a timecode and source
- Editors still own the cut
What was failing
A non-fiction house sat on hundreds of hours of interview media per series. Producers burned days hunting for a usable beat, “when does she talk about the supply chain?” Full transcripts pushed into a public chat model produced loose answers, moved rights-sensitive material off-network, and left no frame-level source. Leadership refused automated picture edits.
What the team built
We built a private media library: speech-to-text with speaker labels, search across transcript and metadata, and candidate moments with in/out points and sources. Editors approve what enters assembly. Generation stays inside the production network. Nothing writes to the timeline without a person.
What operators got
Research and assembly prep moved from multi-day scrub marathons to search you can open at the source. Candidates arrive with timecode; senior editors still cut the picture. Sensitive media never left the controlled path. Cost fell versus re-feeding full transcripts to a public model on every query.
Where AI helped
Transcribe, search, and propose candidates with evidence. Does not auto-cut or publish.
Where software carried it
Ingest, index, search, permissions, approvals, audit log, export handoff.
- 03
Major music label · royalties
Walk every payment line back to the contract clause that justified it
16 weeks across the catalog
- One operational path for the catalog
- High-90s% split accuracy on checked samples
- Human approval before money moves
What was failing
Split terms lived in contracts; finance still lived in spreadsheets. New releases outran re-keying. Auditors could not walk a payment line back to a clause. Letting a model invent rates was never on the table.
What the team built
Contracts entered a database with page-level sources. Models proposed split tables; legal and finance approved before funds moved. Ordinary software did the arithmetic. Everything material was logged.
What operators got
One path for the catalog; new releases join the same flow. Standard closes moved from multi-week to same-week on the measured set. Accuracy held in the high nineties on sampled checks, and the trail from payment line to clause still opens.
Where AI helped
Read contracts, propose tables, flag anomalies. Does not invent a rate or release a payment.
Where software carried it
Calculation engine, catalog binding, reconciliation, approvals, audit log.
Figures are approximate and anonymized. Under NDA we can walk through how they were measured.
Next step
If it has to run in production, and still make sense years from now, talk with the team.
Tell us what is breaking, or what must not break, in a few lines. We will say whether we are a fit, what a first real version looks like, or whether a simpler path is wiser.
Custom build. Shared path: write the bar, ship a first real version, measure, hand over.