// NIJITECH · CORE TECHNOLOGY
Model orchestration is the real work of taking AI to production — not just sending a request to a model. The real work is deciding which model runs, verifying the output and being able to show where the result came from. niji core is the name of that layer.
niji core is the shared AI backbone that nijitech products and customer solutions run on. It converts input into a common representation, decides which model runs based on the type of task (model orchestration), verifies the output it produces and ties it to its source. Data is processed in the European Union region; critical decisions pass human approval.
00 — THE CONCEPT
Model orchestration is the management of several AI models on a single pipeline. The system decides which model runs, based on the type of input and the requirements of the task, processes the result and produces one consistent output.
Sorting a text into one of two categories and running a multi-step analysis do not demand the same capability. Putting the strongest model on every job is like renting an excavator to drive a nail: it works, but the cost makes no sense.
The cost of a model call is small per transaction, but it multiplies with volume. In a system making ten thousand calls a day, routing half of the tasks to a smaller model means a visibly lower bill and visibly lower latency.
A system built on top of one model provider stays exposed to that provider’s price changes, quota limits and policy updates. The orchestration layer loosens that tie: the model changes, the pipeline stays the same.
The best model today can be in second place six months from now. Changing a model call buried in application code is a development job every time; the same change in the routing layer is a configuration job.
01 — HOW IT WORKS
The pipeline runs in the same order in every product; what changes is what the input is and what the output turns into.
Audio, text and system signals — three formats, one stream.
The audio of a meeting, a customer’s WhatsApp message and a marketplace price feed look nothing alike. The input layer converts them into a shared representation: audio becomes text, free text gets structured, system signals queue up with a timestamp. The layers that follow do not need to know where any of it came from.
Which task goes to which model — the system decides.
Sending every task to the strongest model is expensive; sending every task to the cheapest one produces wrong answers. The routing layer looks at the type and the requirements of the task and decides which model runs. A small model is enough for a short classification, while a more capable one steps in for multi-step reasoning. That decision is not fixed; it changes with measurement results.
Model output is never used raw; it passes a check first.
The model produces a result, but that result does not reach the user raw. Does the output match the expected format, is it consistent with its source, is it above the confidence threshold — all of this is checked. Output below the threshold is either regenerated or sent for human approval. The real defence against hallucination sits in this layer.
A summary, a score, a match — and the source of each.
The output depends on the product: a meeting summary, a profitability score, a driver match. What does not change is that every output is tied to its source. Which sentence a summary came from, which data a score derives from — all of it is recorded. Without that link neither audit nor appeal is possible.
02 — FOUR PRINCIPLES
All four are choices; each has a cost and a reason. We have written down both below.
With an on-premise deployment the data stays inside your own infrastructure and does not leave it. In cloud use, processing takes place in the US and EU regions; KVKK Art. 9 safeguards apply to cross-border transfers. The full list of sub-processors is in Exhibit C of the Data Processing Addendum.
Article 9 of the Turkish data protection law (KVKK) makes cross-border transfers subject to consent and safeguards; the GDPR treats data residency as a basic condition of auditability. An on-premise installation satisfies both without entering a transfer debate at all: the data never leaves the organisation. In cloud use, processing takes place in the US and EU regions and Article 9 safeguards apply to the transfer.
Model calls go through infrastructure on-premise or in the cloud. When the buyer’s legal team asks “where does the data go”, the answer is one sentence.
Every result is tied to the source it came from.
You cannot appeal a black-box output. If you cannot show why the result is the result, you cannot debug it, you cannot audit it, and the user will not trust the system. Explainability is an architectural choice, not a reporting feature.
Every item in a meeting summary traces back to a sentence in the transcript, a product score to the demand and competition data behind it, a budget shift to the performance change that triggered it.
Critical decisions pass human approval before they go out.
Fully autonomous decisions make sense when the cost of a mistake is low. For actions that are expensive to undo — issuing an invoice, sending a customer reply, shifting budget — the approval step is not a shortcoming of the system but its safeguard.
Which decisions require approval is set during design — a threshold planned from the start, not a control bolted on later.
Processing runs close to the user.
In a realtime system, latency matters as much as accuracy. A summary that arrives while the meeting is still running becomes useless two seconds late. Every request that travels to a distant region and back eats into that budget.
Light operations run close to the user; heavy inference is left to the central layer. That split lowers both latency and cost.
03 — ENTERPRISE COMPLIANCE
Compliance is an architecture job, not a paperwork job. The table below shows which legal requirement is met by which design decision.
| Requirement | Basis | Architectural answer |
|---|---|---|
| Data must not leave the country | KVKK art. 9 | Processing takes place on-premise or in the cloud. |
| Automated decisions must be contestable | GDPR art. 22 | Human approval on critical decisions; every output tied to its source. |
| Processing must be auditable | KVKK art. 12 · GDPR art. 30 | The input, routing decision and output chain is logged and traceable. |
| Data minimisation | KVKK art. 4 · GDPR art. 5 | Retention periods are defined in the contract; indefinite storage is not the default. |
| Data must not be used for other purposes | Purpose limitation | No by default. There are two exceptions: in proof-of-concept work, where the participating company has given written permission; and for the development of a customer's own platform, at that customer's own request. Outside these cases customer data is not used for model training, and a model trained on one customer's data does not serve another. |
04 — WHAT RUNS ON IT
Whether an architecture works only becomes clear when it is tried on different problems. The sixteen products below do very different things, yet they run on the same pipeline.
nijibotAudio → transcript → summary → action
WhatsModMany channels → one inbox → approved reply
AMZAIPlusListing → price → Buy Box
Amz CalculusDemand + competition + cost → profitability
cep taxiLocation + demand → match → payment
OtoWosInstalments + escrow + legal archive
AdMostPerformance → budget allocation → ROAS
PylunEVM-compatible chain + DeFi layer
CMDeskRemote access → SSH → bulk deployment
KovanlyVirtual machines → networking → backup
LattixPeer-to-peer network → end-to-end encryption
HiriCandidate pool → scoring → human approval
Global Company ServiceDocuments → incorporation → banking and tax registration
Tera LIVELive stream → voice room → gift economy
Exporia.ioCatalogue → buyer list → outreach sequence
Yeşil DünyaMeasurement → source attribution → reduction trackingTERMS
Rather than repeating the definitions here, we keep them in one place in the glossary.
FREQUENTLY ASKED QUESTIONS
niji core is the shared AI backbone that nijitech products and customer solutions run on. It takes the input, decides which model runs, verifies the output and ties it back to its source. All products use the same core.
In practice they describe the same idea: managing several models on one pipeline. "Model orchestration" emphasises the decision of which model runs; "AI orchestration" emphasises the whole of models, tools and data flows. niji core covers both layers.
We are not tied to a single model. Different models step in depending on the type of task, and that choice can change with measurement results. Model independence is a deliberate decision: an architecture dependent on one provider stays exposed to that provider’s price and policy changes.
There is a three-layer defence. The output is checked against the expected format; its consistency with the source is verified; results below the confidence threshold are either regenerated or sent for human approval. On top of that, tying every output to a source makes a fabricated result visible.
The default approach is retrieval, not training: we let the model reach your data at query time (RAG). Because the data is not baked into the model, it stays current and it remains possible to withdraw it. Cases that genuinely need custom training are assessed separately.
Accuracy is measured per task, and every project builds its own evaluation set. A fixed “x percent accurate” claim is meaningless without saying on which task and on which data it was measured. We agree the metric together during discovery.
That is what the integration layer is for. We build working connections to calendar, CRM, messaging and accounting systems. Which system you connect to — and whether it is technically possible — becomes clear during discovery.
For general questions see the FAQ page · for our services see enterprise solutions
GET STARTED
We work out together in a discovery call which of your work fits this pipeline.
Book a discovery call →We would like to use Google Analytics to understand how the site is used. It sets cookies and transfers data to the United States. The site works exactly the same if you decline.
