// NIJITECH · GLOSSARY
This page defines the terms that appear in nijitech's product and technology language, explainable AI among them. Each definition is written to stand on its own — quotable and linkable.
Model orchestration is the management of multiple 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 coherent output. Because each task is routed to the most suitable model rather than being locked to a single one, it offers advantages in both cost and accuracy.
At nijitech this layer is called niji core; all products run on this backbone.
#Explainable AI is the ability to show, in human-understandable terms, why an AI system produced a given result. The output stops being a black box: which input it came from and on what grounds can be traced. Auditing, appeal and debugging are only possible with this traceability.
In practice it means each item in a meeting summary can be traced to its source in the transcript, and a product score to the demand and competition data behind it.
#An AI agent is a system that plans its own steps toward a given goal, calls tools, and adjusts course based on intermediate results. What separates it from a model that answers a single question is its ability to carry a multi-step task from start to finish.
An advertising agent, for example, may monitor performance and propose shifting budget, requesting human approval at the critical step.
#A human-approved decision is one where the AI's result is reviewed and approved by a person before it is applied. The AI prepares the recommendation; the person accountable makes the call. It is a standard safeguard for actions that are irreversible or costly.
In nijitech products, steps such as budget shifts, payments and outbound communication pass through this gate.
#EU data residency means personal data is processed and stored only on infrastructure inside the European Union. Data is not moved outside the EU for processing. It is a decisive choice under GDPR's transfer provisions and comparable KVKK regulations.
For organisations in Türkiye, the practical value is being able to evidence where data was processed during a compliance audit. In nijitech products this depends on the deployment model: with an on-premise installation the data stays in the customer own infrastructure; in cloud use, processing takes place in the US and EU regions.
#Edge scaling is running computation on nodes geographically close to the user rather than on a central server. Because data travels a shorter distance, latency drops and load does not concentrate in one place. It is decisive for work that requires real-time response.
Used in scenarios where seconds matter, such as live transcription and location-based matching.
#RAG is the practice of having a language model search for and retrieve relevant documents before answering, then grounding the answer in those documents. Instead of relying on what it memorised during training, the model consults current, verifiable sources. This lowers the risk of fabrication and makes the output traceable to a source.
It is one of the technical foundations of explainability: the document an answer came from can be shown.
#A hallucination is when an AI model confidently produces information that does not exist. The answer is linguistically coherent but factually wrong. It is one of the core risks in enterprise use.
Source-grounded generation (RAG), human approval and evaluation sets are used together to reduce this risk.
#An evaluation set is a body of test examples with known expected answers, used to measure the accuracy of an AI system. When a model or prompt changes, this set is what tells you whether the system improved. A model change made without measurement is guesswork.
Combined with bias-aware metrics, it lets you track not just average performance but performance across groups.
#The Buy Box is the seller position tied to the "Add to Cart" button on an Amazon product page. Where several sellers offer the same product, Amazon's algorithm picks one, and the majority of sales go to that seller. Price, stock, shipping speed and seller performance all influence the choice.
Managing Buy Box competition is one of the most direct levers on sales volume on Amazon.
#ROAS is the ratio of revenue generated by advertising to the amount spent. A ROAS of 4 means 4 units of revenue for every 1 unit spent. Because it condenses campaign performance into a single number, it is the core metric of advertising optimisation.
Autonomous decisions about how to distribute budget across campaigns are usually made against this metric.
#Escrow is holding a buyer's payment with a trusted third party rather than passing it straight to the seller, releasing it only once the conditions of the transaction are met. It protects the buyer against paying without receiving, and the seller against delivering without being paid.
Used in high-value transactions conducted remotely, such as vehicle sales.
#A peer-to-peer network is a model in which devices connect directly to each other without passing through a central server. In the classic client–server model all traffic flows through a server in the middle; in the peer-to-peer model that intermediate stop does not exist. The consequence is threefold: latency drops because the path is shorter, shared server capacity does not form a ceiling, and there is no single point of failure whose outage stops the connection.
Among nijitech products, Lattix uses this model.
#Virtualisation is the division of a single physical server into several independent virtual machines. Each virtual machine runs its own operating system and stays isolated from the others. Containers, by contrast, share the operating system kernel and provide lighter isolation: higher density on the same hardware, but a thinner isolation boundary than a virtual machine.
Among nijitech products, Kovanly manages virtual machines and containers in one panel.
#End-to-end encryption is a form of encryption in which data can only be decrypted at the sending and receiving ends. No component in between — network provider, server, or the service itself — can read the content. That is what separates it from transport encryption: with transport encryption the data is decrypted when it reaches the server; with end-to-end encryption it is not.
In remote access and networking products this distinction is decisive: it determines whether the service itself can see the content.
#A guardrail is a predefined boundary an autonomous system cannot cross. Inside the boundary the system decides freely; when it reaches the edge it stops or asks a human. A daily spend ceiling in advertising, a transaction limit in a payment system, and the authority limit of a tool in an agent architecture are all guardrails.
In autonomous systems a guardrail is the basic way of bounding risk without giving up speed.
#EVM compatibility means a blockchain can execute Ethereum Virtual Machine bytecode. In practical terms: smart contracts written in Solidity, wallets such as MetaMask, and development tools such as Hardhat and ethers.js work on that chain as they are. The team does not have to learn a new language or toolchain.
Among nijitech products, Pylun is an EVM-compatible chain.
#Finality is the point at which a blockchain transaction becomes irreversible. Under probabilistic finality the chance of reversal falls as blocks are added on top but never reaches zero, so several blocks are awaited. Under instant finality the transaction settles in a single block and cannot be reverted. In payment and commerce applications this difference translates directly into waiting time.
The consensus algorithm determines which kind of finality a chain offers.
#An on-premise deployment runs the software on the organisation’s own servers rather than in the vendor’s cloud. The data never leaves the organisation’s infrastructure, so the question of cross-border transfer never arises. Unlike cloud deployment, hardware, updates and backups remain the organisation’s responsibility.
nijitech products support both on-premise and cloud deployment; the choice depends on data sensitivity and operational capacity.
#An MCC is the four-digit code that card schemes assign to each merchant according to the business it does. Banks assess risk by this code: in some categories the number of instalments is capped or switched off entirely. High-ticket categories such as car dealerships, jewellery and fuel run into this limit most often.
When a dealership cannot offer instalments, the cause is usually not the customer’s limit but the merchant’s MCC.
#A chargeback is a cardholder asking their bank to reverse a payment that has already been made. If the bank upholds the claim, the amount is pulled back from the merchant — even if the goods were delivered. In high-value, non-face-to-face sales this is the merchant’s single biggest collection risk.
In an escrow structure the money is held until the transaction completes, so the chargeback risk does not sit with the merchant.
#Fine-tuning means retraining a ready-made AI model on an organisation’s own examples so that its behaviour changes permanently. The model learns tone, format and domain vocabulary. The difference from RAG: RAG fetches knowledge from outside at run time, while fine-tuning writes it into the model itself.
The two are not rivals: RAG solves freshness, fine-tuning solves tone, and most enterprise deployments use both.
#An ATS is software that runs the hiring process in one place, from job posting to offer. It collects applications, moves candidates through stages and keeps a record of correspondence. A classic ATS is an archive: it holds the data, but a human does the evaluating.
With a talent-intelligence layer the system stops being an archive: it scores candidates on the same criteria and runs part of the interview itself.
#e-Residency is a state-issued digital identity granted by Estonia. It lets the holder set up an Estonian company and sign banking and accounting transactions remotely, without ever visiting the country. It is not a residence permit or citizenship — only an authority to transact digitally.
Because Estonia does not tax undistributed profit, it is favoured by digital businesses that keep earnings inside the company.
#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.
