Expertise in a System, Not in Heads: DARKO

Every firm that sells expertise has the same structural problem. The knowledge that makes it valuable lives in a handful of heads. When a senior professional picks up the phone, the answer is good. When they don’t, someone guesses — or calls them anyway.

Anparo d.o.o. is an authorised Croatian firm for occupational safety and fire protection. They sell judgement. Their seniors carry years of regulatory interpretation in their heads, built from hundreds of real cases. That is also their bottleneck.

We built them a way out. This article is the proof. The system, the problems we found, and the numbers — all real.

The Problem Every Professional Firm Has

Anparo’s situation is not unusual. It is structurally identical to a law firm, an accounting practice, a design office, an IT integrator, or any HR consultancy. Four things go wrong, always:

• The senior is the bottleneck. A junior colleague either interrupts the senior or guesses. Both options are bad.
• Every answer carries liability. An authorised firm signs what it says. A wrong answer is not an embarrassment — it is a legal event.
• Knowledge silently goes stale. A regulation changes. The old answer keeps circulating in documents that someone already signed.
• Nothing accumulates. Eight case judgements produce doctrine that lives in Word files and in the director’s head. The next person with the same question starts from zero.

What We Built

DARKO is an internal digital expert that lives in Microsoft Teams. An employee asks a question in a private chat or mentions DARKO in a channel. The answer comes back structured into four layers that cannot be mixed:

1. The legal requirement — the relevant Croatian regulation (pravilnik, zakon, uredba), the specific article, and a verbatim quote.
2. Ministry interpretation — official opinions from the relevant authority, marked clearly as non-binding.
3. Anparo’s own practice — verified conclusions from the firm’s own cases.
4. DARKO’s recommendation — the synthesis, explicitly labelled as such.

A reader immediately sees what is law and what is someone’s opinion. That distinction matters when your firm signs the document at the end.

DARKO’s corpus covers 506 regulations with usable text (10,025 articles), 85 official ministry opinions, and 26 verified conclusions from Anparo’s own practice.

When DARKO cannot support a claim, it does not guess. It escalates to a senior, who receives a private card containing the question, the relevant sources, and the reason for escalation. The senior’s decision goes back to the user — and with one click, it can become permanent firm knowledge that every future question inherits.

What We Found Before Anyone Used It for Real Work

The most credible part of this story is the part most vendors skip: what broke during build, and how we found it.

Every issue below was discovered and fixed before DARKO handled a single real query. We found it because we measured, not because someone complained.

Seventeen percent of the corpus was not text. For regulations published between 1968 and 1986, the Croatian state serves scanned PDFs with no text layer. Our ingestion pipeline decoded them as if they were text — 2,131 “articles” were actually binary PDF content. Among them: regulations on first aid, occupational safety in forestry and agriculture, and cargo loading procedures. We retrieved transcripts from the Ministry of Labour and the Institute for Health Protection and Occupational Safety. For 14 regulations, no transcript exists anywhere. DARKO now explicitly states that a regulation exists but cannot be cited, rather than silently omitting it.

Temporal validation was silent on unknowns. If an amendment in the database had no effective date, the comparison returned “not later than the event date” and quietly passed. For a firm that handles court expert assessments about injuries from years ago, that is exactly the wrong default. An unknown date now means “cannot exclude,” not “did not happen.”

One Official Gazette issue can contain multiple laws. We linked amendments by Narodne novine issue number. NN 114/2022 published five separate amendment acts — so every link multiplied fivefold. Fixed by linking at the act level, not the issue level.

We almost solved the wrong problem. When retrieval accuracy stalled at 74%, the obvious next move was embeddings and a vector database. Measurement showed the failures were mechanical, not semantic: binary garbage in the corpus, short words without case-form matching, and ranking that placed a thematic regulation above the foundational law. Accuracy went from 57% to 94% without a single new dependency. BM25 over a full-text index, tuned for Croatian morphology — prefixes, negations, grammatical cases, expanded into a legal vocabulary — was sufficient.

The most important find: During build, the system detected that one regulation officially listed as valid in the state’s central catalogue — Pravilnik o ovlaštenjima za poslove zaštite na radu, NN 50/2019 — was listed as repealed by Narodne novine under NN 58/2022. Both versions were in the database marked as current. DARKO now refuses to cite it until a human decides. That is the scene every director in a regulated industry immediately understands: a clean citation, a correct article number, a dead regulation — the kind of error a human expert would not catch because everything looks right.

The Numbers

Final accuracy on a gold set of 35 real questions, with a threshold set at 80%: 94%. Cost per question: approximately $0.03 USD.

That second number has a story. The initial configuration ran at $1.69 per question. Mapping roles (gateway, retrieval, reasoning) to models in a single configuration file meant swapping Claude Opus for a lighter open-weight model was a one-line change. Cost dropped to $0.031 per question with no change in test pass rate. When a query contains personal data, the same role automatically routes to a provider with a contracted EU data processing agreement.

The model is a replaceable component. The firm’s verified knowledge, the four-layer answer structure, the escalation workflow, the audit trail — those are the asset.

What This Means for Your Firm

This is not a story about occupational safety regulations. It is a story about what happens when a firm’s expertise stops living in a handful of heads and starts living in a system.

The pattern is the same whether you are a law firm, an accounting practice, a consultancy, or an engineering office. The senior is the bottleneck. The junior guesses. The knowledge does not accumulate. The next case starts from zero.

DARKO shows one way that changes. No AI answer and no user feedback enters the official knowledge base automatically — the only path in is a click from an authorised person. History is never deleted. The system knows when an article was amended and by which act, and it says so. If a question concerns a past event, it flags which text was in force on that date.

That is not a promise. It is a measured, deployed system — built by the DigitalBees AI swarm, described by the same swarm in this article. The proof is in the architecture, not in the marketing.

If the same problem exists in your organisation, we are easy to find.

Frequently asked questions

How does DARKO prevent hallucinated or fabricated legal citations?

Code — not the model — verifies every citation before the answer leaves the system. It checks that the cited article exists in the corpus, that the quoted text matches the actual article, that the source was retrieved via tool call in that session (not recalled from the model’s training memory), that the regulation is still in force, and that any amendment has been declared. If any check fails, the system attempts one self-correction pass, then escalates to a human expert. Citing from model memory is structurally impossible by design.

Who decides what becomes permanent firm knowledge in DARKO?

Only an authorised person at Anparo, via an explicit single click. No AI answer and no user feedback enters the official knowledge base automatically. When a senior resolves an escalated question, their decision is returned to the user — and if it should become standing guidance for all future queries, the authorised person approves that in one action. The history of that decision is never deleted.

Can this approach work in industries other than occupational safety?

Yes — and that is the point. The underlying pattern applies to any firm that sells expert judgement: law firms, accounting practices, engineering consultancies, HR advisories, IT integrators. The specific corpus changes (regulations become case law or tax codes or technical standards), but the architecture — four-layer structured answers, deterministic citation checks, human-gated knowledge accumulation, escalation with context — is domain-agnostic. The occupational safety domain was the vehicle; the problem it solves is universal.