Skip to content
TrinolIT
TrinolIT

We build thesystems thathave to be right.

Software, AI and computer vision engineering out of Dhaka, for teams who have to defend the answer. Every system we ship can show how it reached one — in English and in বাংলা.

Nexora AI — point of sale

One checkout. Six writes. No middle state.

A sale touches inventory, cash, tax and the ledger at the same instant. Systems that write those one at a time are systems whose books are wrong and nobody knows yet.

When the fourth write fails, the first three are undone. The database is the thing enforcing that, not the application.

  1. Saleline items, totals, tax basis
  2. Stock movementledger entry, balance lock
  3. Paymenttender, change, settlement
  4. VAT recordrate, base, output tax
  5. Journal entriesdebit, credit, immutable
  6. Audit eventactor, hash, chained

commit — all six, atomically

What we build

Six things we do

Built to be inspected.

Each of these is a capability we can point at working code for, or state plainly that we cannot yet.

01Systems, not features

Product and platform engineering

Full systems rather than features: transactional backends, multi-tenant data models, an interface on top, and the deployment and release pipeline underneath.

What this involves

02Scoped and grounded

AI systems that stay inside their limits

Retrieval scoped to the data the user is allowed to see, tools the model selects from rather than composes, and a grounding check that every figure it states came from a real result.

What this involves

03Correct under load

Data integrity and correctness

Money as decimals end to end, immutable ledgers, reversible migrations, and concurrency handled where it actually bites — the last unit in stock, the double-submitted payment.

What this involves

04Isolation, tested

Security and tenancy

Isolation defended in layers and tested adversarially, tamper-evident audit trails, encryption of the fields that need it, and responses that do not leak the existence of what they refuse.

What this involves

05Per class, not headline

Machine learning and computer vision

Forecasting, anomaly detection and detection pipelines built against a named benchmark, reported per class rather than behind a single headline score.

What this involves

06Bangla on its own terms

Bilingual and Bangla-first systems

Bangla and English handled on their own terms rather than English-first with a translation layer — tokenisation, code-switched input, and OCR on scanned documents.

What this involves

The work, andhow to check it.

Shipped — Business software

Nexora AI

A multi-tenant ERP, POS and business intelligence platform where a checkout either writes the sale, the stock movement, the payment, the VAT record, the journal entries and the audit event — or writes none of them.

  • FastAPI
  • PostgreSQL 16
  • SQLAlchemy 2
  • Next.js
  • TypeScript
  • Celery
  • Qdrant
  • Docker
The Nexora AI command centre, labelled in English and Bangla: sales efficiency, cash position and stock alerts across the top, then revenue, gross profit, receivables, payables and inventory value, with a weekly sales chart and a ranked product list.
Nexora AI, in English and বাংলা
12

Delivery phases complete, each with a written exit state

docs/ROADMAP.md in the public repository, with CI green on main

99

Test files across backend and frontend

Counted in the public repository

80%

Backend coverage floor

Enforced by `make test`; the build fails below it

26

Database migrations, all reversible

alembic/versions in the public repository

In build

Student GPA and management system

Grade recording, GPA calculation and student administration for an education provider. In build now.

No results published yet. They appear with the method that produced them.

TrinolIT Labs — try it

Read someoneelse’s invoice.

Pick who you are and which record you want. Isolation in Nexora AI is three independent layers, and a cross-tenant read returns 404 rather than 403 — because 403 would confirm the record exists.

This runs in your browser and touches nothing. It models the rule; the code that enforces it is in the public repository.

You are signed in as
You request
  1. Layer 1Derived tenant contextSession resolves to t_ashal. Request carries no tenant of its own.
  2. Layer 2Automatic query filterScoped query matches the row.
  3. Layer 3Row-level securityPolicy permits. The row is returned.

Response200 OKInvoice 8841 belongs to your tenant, so you get it.

How we work

Five steps. One of them is the reason to hire us.

  1. Scope

    We work out what the system has to get right and what it is allowed to get wrong, and write it down before anyone opens an editor.

  2. Decide

    The decisions that are expensive to reverse get recorded with their cost, not just their benefit. A decision record that lists no downside has not been thought through.

  3. Build

    Production code from the first commit: typed, migrated, tested against real infrastructure, in your repository rather than ours.

  4. Attack it

    The isolation, the concurrency and the money paths get an adversarial test suite whose job is to break them. This is the step that separates a system that works from a system that has been shown to work — and it is the one most agencies do not have, which is why they cannot tell you whether their software still holds.

  5. Hand over

    Documentation, decision records, the test suite, the deployment, and a walkthrough with the people who will own it. You should not need us afterwards.

Start here

You havesomething thathas to be right.

Tell us what it has to get right and what it is allowed to get wrong. If it is not something we should take on, we will say so and point you somewhere better.

Reply time
We reply within one working day.
Or write directly
zasif855@gmail.com
Where
Dhaka, Bangladesh