doItWisely
Evidence-backed decision analysis for regulated domains – deterministic, traceable, without an LLM in the core path.
A decision-support engine rather than a chatbot: the structured DecisionAnalysis as the source of truth, declarative rules, a knowledge graph, provenance and traceable reasoning paths – opened up in production for domains such as CE certification, German GmbH law and battery know-how.
- Technical challenge
- Regulatory and business decisions – for example around CE certification, German GmbH law or battery know-how – were to be made reproducibly and backed by evidence: traceable, without pseudo-precision and without a language model replacing subject-matter truth.
- Architecture
- Modular monolith: Fastify API and React frontend; PostgreSQL + pgvector for sources, rules, parameters, company data, gold cases and audit; Neo4j Community for a two-layer knowledge graph (semantic knowledge plus a rebuildable rule-derived graph). Declarative JSON rules with a typed condition AST, a deterministic calculation engine, a source-authority model and a strictly downstream, provider-neutral LLM.
- Engineering challenges
- Separating semantic knowledge from rule-derived effects, first-class provenance and temporality, conflicting sources treated as insight rather than noise, money arithmetic in minor units, qualitative assessment without invented scores, as well as operating two databases.
- Demonstrated capabilities
- Knowledge-graph engineering, rule-based and explainable decision systems, deterministic calculation, evidence and provenance modelling, a multilingual RAG architecture and a validated decision test suite.
doItWisely is a Riegel Systems engineering case for evidence-backed decision analysis in regulated domains. It is neither legal advice nor a chatbot, but a decision-support engine: the structured DecisionAnalysis is the truth, and text is merely a projection of it.
doItWisely has been in production since today. Domains it has learned include CE certification, German GmbH law and battery know-how – each as verified knowledge in the knowledge graph, not as gut feeling or a plausible LLM answer.
Learned domains
- CE certification
- German GmbH law
- Battery know-how
Decision path
- Question → Intent & company context
- Knowledge Graph (Neo4j)
- Applicable rules (declarative)
- Deterministic calculations
- Scenario comparison
- DecisionAnalysis
- Evidence enrichment (RAG)
- LLM formulation (downstream)
Architectural points
- Structure is the source of truth – text is projected from the result, not the other way around
- Two-layer knowledge graph: authoritative semantic knowledge plus a rebuildable, rule-derived effect graph
- Declarative rules with a typed condition AST instead of hidden subject-matter logic in code
- Provenance and temporality are first-class and machine-readable
- Conflicting sources are detected, preserved and reported – authority drives ranking and confidence, not the subject-matter truth automatically
- RAG strictly downstream: verified knowledge before runtime document interpretation
- Money in minor units, no floating-point arithmetic for amounts
Key takeaway
doItWisely stands for the ability to build complex domains as explainable, deterministic and auditable systems – with a knowledge graph, a rule set and complete evidence instead of plausible but unverifiable LLM answers.
Technologies & concepts
A similar engineering project?
If your project needs comparable technical depth, we discuss feasibility, architecture and effort – concretely rather than in pitch format.