Applied AI · Developer tools

Rikaha

A software-system clarity and diagnostic reasoning platform for people who can build with modern tools faster than they can reliably understand or debug the systems those tools produce.

Applied AITypeScriptNext.jsFastifyPostgreSQLEvaluationCI/CD

The problem

Generated software is easy. A reliable mental model is not.

Rikaha grew from a problem I repeatedly hit while learning and building: AI, frameworks and deployment tools can produce applications, containers, reverse proxies and database configuration faster than a developer can form a trustworthy picture of how all the pieces actually interact.

The useful product is therefore not “classify this error.” It is a system that reconstructs context, distinguishes evidence from assumptions, chooses the next high-information check, interprets the returned evidence and keeps going until the cause is isolated and the fix is verified.

Reasoning model

Troubleshooting as an evidence loop, not a one-shot answer.

01Problem + context
→
02Observed / assumed / unknown
→
03Rank hypotheses
→
04Next high-information check
→
05Evidence update
→
06Cause · fix · verify

The product direction is deliberately progressive: uncertainty can shrink as evidence arrives instead of being hidden behind one confident first response.

Implemented / accepted evidence

Product, data and research boundaries are designed to survive change.

Product runtime

Next.js + Fastify

A thin web client and standalone API keep product semantics out of page components and preserve future client replaceability.

Durable state

PostgreSQL

Runtime and migration authority are separated; the development path has been proven with TLS verify-full, SCRAM and role-scoped access.

Diagnostic model

Progressive state

The target loop separates observed facts, assumptions and unknowns, then updates hypotheses as new evidence arrives.

Research boundary

Separate Lab

Benchmarks, evaluators, engine evolution and falsification live outside the product repository and cannot silently become production truth.

Engine lineage

Immutable

The first modern accepted engine has a fixed identity, manifest and scientific decision; later behavior changes require a new version.

Trial boundary

Fail-closed

Scientific acceptance does not equal product promotion. The owner-visible trial path is separately gated and preserves exact engine/run provenance.

System boundary

Research can recommend. Product decides what ships.

PRODUCT / RUNTIMERikaha Platformweb · API · product state · evidence · release
reviewed handoff↔trial evidence
RESEARCH / EVALUATIONRikaha Intelligence Labresearch · benchmark · evaluator · engine lineage

First modern engine

rk-baseline-v1.0.0

Scientific statusAccepted
Product-visible trialSeparately gated

A useful architectural reset

The project improved by admitting when the abstraction was wrong.

01

Earlier lesson

Scenario matching became too close to the definition of intelligence.

Current direction

Categories remain useful for retrieval and organization, but the core product is a progressive diagnostic episode.

02

Earlier lesson

Moving routing rules into database tables improved cleanliness but regressed measured matching quality.

Current direction

PostgreSQL stores durable truth and lineage; it is not used as a substitute for a reasoning model.

03

Earlier lesson

Product UI, matcher research, thesis work, mobile and deployment evolved together.

Current direction

Product/runtime and research/engine evolution now have separate repositories, contracts and acceptance gates.

Current product slice

Episodes preserve evidence and outcomes before intelligence is allowed to own truth.

The current product work models a diagnostic episode as durable user problem context, append-only evidence and user-confirmed outcomes. The engine-neutral trial lane adds immutable engine-run input/output/provenance receipts.

A deliberate correction removed engine-generated observations and hypotheses from canonical mutable product truth before the schema was promoted. The platform can display an engine's interpretation without pretending that one engine's vocabulary is the permanent data model.

What I owned

Product direction, architecture and the evidence required to trust changes.

  • Product problem and target diagnostic loop
  • Greenfield architecture and repository boundaries
  • Database/runtime/release constraints
  • Research-to-product promotion rules
  • Engine trial acceptance criteria
  • Owner-visible product and usefulness judgment

AI-assisted workflow

ChatGPT and Codex are major implementation and research collaborators. My responsibility is to decide what the product is, separate evidence from assumptions, choose architecture and constraints, reject weak abstractions and accept changes only when source, CI, database/runtime and product evidence agree.

Honest boundary

Accepted research is not the same thing as a finished public AI product.

Rikaha demonstrates applied AI product architecture, diagnostic-state design, benchmark/evaluator discipline, immutable engine lineage, PostgreSQL-backed product state and research-to-product governance.

It does not claim universal debugging, autonomous production remediation, broad unseen generalization, commercial scale or foundation-model research. The controlled owner-visible trial is intentionally a separate acceptance step.

Another AI systems case study

ROS — a persistent local-first assistant with a separate GPU inference plane.

View ROS case study