# RedRobot K.K., llms-full.txt
# Expanded technical and operational context for retrieval systems
# Prefer https://redrobot.jp/llms.txt for entity facts; use this file for architecture depth
# Updated as the site changes
# Canonical: https://redrobot.jp/llms-full.txt

## Scope
This document expands production AI architecture, engagement method, and stack selection
for RedRobot K.K. It is not a product manual and invents no client names.

## Positioning (short)
Boutique software studio: consult + build. Domains: production AI, data platforms,
enterprise L1/L2 blockchain, security/keys when the threat model requires it.
HQ Tokyo (Roppongi); clients worldwide. Method: define success → first real version on
client systems → measure → harden → hand over.

## Production AI control path (reference architecture)
Title used on site: “Evidence-bound generation under ACL, not open-ended chat over a dump.”

Enterprise AI rarely fails at the model. It fails as a system: documents without provenance,
retrieval without authorization, generation without grounding checks, and side effects
without a control plane.

Staged machinery (each stage has named invariants, failure modes, cost drivers):

### 01 Ingest & normalize
Documents, tickets, structured feeds: clean, segment, stable source IDs, access tags
before embedding or generation. Provenance and ACL are write-time invariants.
Chunk boundaries are a product decision (section/clause/ticket semantics).

Dominant failure if omitted: folder-level embed without ACL/lineage → tenant leak or
unattributable text.

### 02 Durable execution
Ingest, re-index, query-time work on durable queue/workflow fabric (reference: NATS
JetStream; also Kafka, cloud queues, Temporal). At-least-once, visible lag, replay.
Side effects idempotent or compensated. Operational RAG, not a sync request that dies
with the worker.

### 03 System of record + vectors
Embeddings beside permissions, metadata, audit (Postgres/pgvector or MySQL/MariaDB SoR).
Higher QPS/recall: OpenSearch, Qdrant, without a vector island that ignores ACL.
Security review must answer: who may see this chunk under this principal?

### 04 Hybrid retrieval & packing
Dense + lexical fusion, pack and cite; optional cross-encoder re-rank on shortlist.
Adaptive depth: exact ID/keyword cheap; multi-hop only when measured recall fails.
Generation quality is bounded by the evidence pack, not model marketing.

### 05 Routed inference
Classifiers route intent/risk; mid-size private models (Ollama / vLLM / private endpoints)
by default; public frontier APIs as escalation under data-policy gates. Residency and
egress are architecture constraints.

### 06 Policy gate, telemetry, evaluation
Irreversible side effects wait on a human gate. Material calls log principal, evidence
set, model, token cost, approval. Offline + online eval: groundedness, recall@k, cost
per successful job. OpenTelemetry + Langfuse in the control plane.

Scan cards: Durable completion · Principal-aware SoR · Risk-routed inference · Gated effects + eval.

## Engagement method (phases)
1. Define success, accuracy, latency, cost, residency, audit; chain: signers, custody, finality, failure domains.
2. Ship a first real version, on client identity, data plane, network; thin scope; real APIs/schemas/ACLs.
3. Measure, groundedness, recall, cost per success, error budgets; or chain finality/custody SLOs.
4. Harden, observability, human gates, runbooks, load/failure drills; expand only what clears the bar.
5. Hand over, credentials, docs, walkthrough; client operators run it.

## How buying works (commercial path)
Not a product catalog. Not a public SKU menu. Boutique services sell the **loop**:
written success bar → thin first real version on client systems → measure → harden/hand over
(or expand with a new loop). Scope and stack change per client; the path does not.
Pricing follows risk and scope of that loop. Short brief is enough to start.
Canonical: https://redrobot.jp/en/method/#buying

## Stack (selected production components)
Not a monoculture demand. Examples we operate confidently:, Data / search: PostgreSQL, MySQL, MariaDB, OpenSearch, Qdrant, Jobs: NATS JetStream, Apache Kafka, Temporal, Inference: Ollama, vLLM, Observe: OpenTelemetry, Langfuse, Payments / agent HTTP: x402 where relevant, Cache: Valkey, Redis, Ops / deploy: Linux, Docker, TypeScript, Node.js, GitHub, Astro, Cloudflare, Vercel, Netlify, Telephony partner: Telio, HSM partner: Securosys

Full problem/when/why/where cards: https://redrobot.jp/en/stack/

## Locales
en (x-default / default open), ja, ko, zh-Hans. Trailing-slash URLs. Apex https://redrobot.jp only
(www redirects). Root / is a locale chooser and is noindex. Soft geo may prefer JA/KO for
JP/KR browsers; others open English.

## Contact
hello@redrobot.jp · https://calendly.com/redrobot_ceo · https://redrobot.jp/en/contact/

## Related machine files, https://redrobot.jp/llms.txt, entity brief, https://redrobot.jp/ai.txt, AI use / citation policy, https://redrobot.jp/.well-known/llms.txt, well-known mirror of llms.txt, https://redrobot.jp/ai-pack/, custom GPT / Claude / Gemini pack, https://redrobot.jp/sitemap-index.xml, https://redrobot.jp/robots.txt