// NIJITECH · BLOG
On technology, software and artificial intelligence.
Weekly writing on the practice of taking AI into production and the architecture decisions behind it.
Posts
Technology6 min read
MCC limits at car dealerships: why the bank cuts instalments
When a dealership cannot offer instalments, the cause is usually not the customer’s limit but the merchant category code. What the code does, and how to work around it.
Technology5 min read
Escrow in car sales: when does the money change hands?
In an escrow structure the money goes not to the seller but to a licensed third party, released once the conditions are met. How the flow works, step by step.
Technology6 min read
Finding export buyers: why buying a list does not work
There is a difference between a purchased email list and finding a buyer. One gives you a name; the other shows why that name would buy your product.
Artificial intelligence7 min read
Voice-first hiring: ATS versus talent intelligence
An ATS is a system of record; talent intelligence is a system of assessment. Treat them as the same thing and candidates who could do the job get cut on keyword match.
Technology6 min read
Blockchain finality: when is a transaction really done?
“Confirmed” and “irreversible” are not the same thing. For a system accepting payments, that difference is what decides how many seconds you wait.
Technology6 min read
The costliest mistake in remote access: forgetting to revoke
In remote access, whoever connects holds the same privilege as the person sitting at the machine. The real risk is not the password but not knowing who can still get in.
Software6 min read
Hardcoding prompts: a new form of technical debt
When model prompts get scattered through source code, every text change turns into a release. This is exactly what happened to SQL queries a decade ago.
Artificial intelligence7 min read
RAG or fine-tuning? Which problem wants which
Both get taken as the answer to a sentence that starts “let us teach the model our own data”. They solve two different problems. Three criteria that settle the choice.
Software7 min read
The black box problem in marketplace repricing
If a system that changes your prices automatically cannot say why it changed them, mistakes cannot be found. Nobody notices until the margin has melted. How to build auditable repricing.
Technology7 min read
Where is your data processed? Mapping the whole chain
“Our data is in Europe” usually describes the first link in the chain. The real question is who that link depends on. A practical way to build a processing inventory.
Artificial intelligence7 min read
Measuring hallucination: how to build an eval set
“The model makes things up sometimes” is an observation, not a measurement. You cannot claim to have fixed a problem you cannot measure. A practical way to build an evaluation set.
Technology6 min read
Cloud on your own hardware: when does it make sense?
The cloud is not always the right answer, and neither is your own hardware. The decision is set by four questions, not by a price table.
Artificial intelligence7 min read
Why AI pilots stall before production
The gap between an AI pilot that works and a system that survives in production is not in the model but in the layer built around it. Four concrete breaking points.
Artificial intelligence6 min read
How many models, how much money? Costing AI
Per-transaction cost is small and invisible in a pilot; multiplied by volume it decides the entire invoice. A practical way to measure it and bring it down.
Software6 min read
Where automation ends and autonomy begins
Rule-based automation works as long as the input arrives the same way every time. The moment it does not, it stops. Where the boundary sits and which approach belongs where.