About eqlx · İstanbul
A senior engineer leads your work, from start to handover.
Every engagement is led by Can Demir, founder of eqlx — an AI and machine-learning engineer who has built machine-learning models that real decisions run on since 2012, and language models and agentic systems in production since 2024.
2010–2012
A large Turkish industrial manufacturer
Manufacturing engineer: production planning and process optimization on the factory floor.
2012–2015
A large Turkish bank
Ran a small and medium-sized enterprise (SME) and corporate lending portfolio, then built the credit-risk data layer the bank's models ran on.
2015–2021
The same bank
Senior data scientist for live credit decisions across SME, corporate, and retail lending: application, behavioral, early-warning, collections, pricing, and fraud models under Basel III and IFRS 9 (International Financial Reporting Standard 9).
2021
An international banking group
Led rating models across every portfolio, real-time scoring, and an early-warning system wired into daily credit decisions.
2023
A European IT consultancy
In Amsterdam: risk-based pricing and limit models for European lenders, taken into monitored production.
2024–2026
A fintech he co-founded
Co-founder and Head of Data & AI: AI agents, retrieval systems, and small language models post-trained with fine-tuning and reinforcement learning, in production on clients' own servers.
Why eqlx
Every case, decided the way your rules say.
Equalize the gap. Close the distance between how a decision should be made and how it is made today.
eql is equal: decisions kept equal to your rules. x is your case: every case, and every institution, is different.
In the logo, the ink bar is the rules and the red bar the decision: level means the decision equals the rules.
How we work
We come into your environment, build beside your team, and hand the system over running.
We come in
We work inside your environment, through the accounts and access your team issues.
We build beside your team
Your people see every step, own the test, and learn to run what we build.
We hand it over running
Documentation, runbooks, and training go with it; our access is removed unless support is agreed.
Integration is the base under every kind we build: we build and integrate, or integrate what you already run.
Three starting points, each a separate decision that stays yours: the Decision Blueprint, a validation build, then production, integration, and handover.
What it is like to work with us, and how our access ends:
Principles
Five rules we keep on every engagement.
Start from the decision.
We begin with the desk, its rules, and who signs; the technology follows.
Use the smallest thing that works.
A rule or a calculation wins when it is enough; a model or an agent is added only where the decision needs one.
Your expert writes the test.
Nothing goes live on a task it has not passed on your own cases.
The limit stays yours.
Inside the limits your policy writes, the system acts; everything else goes to a named person.
The model and the data stay yours.
The weights we train for you are yours under the engagement terms (the base model's license still applies), and your data trains no other system. A model on your servers changes only when your team approves it.
Why now
Bank rules keep the systems in Türkiye, and researchers argue small models can do much of the work.
Regulation
Banks must keep their primary and secondary systems in Türkiye.
Our translation of the Turkish original.
BDDK, the Regulation on Banks' Information Systems and Electronic Banking Services, Article 25(1) · Official Gazette 31069, 15 March 2020 · Source
Research
Researchers at NVIDIA and Georgia Tech argue that small language models are “sufficiently powerful, inherently more suitable, and necessarily more economical” for many of the calls in agentic systems.
The authors' position, not a measured result.
Belcak and others, “Small Language Models are the Future of Agentic AI”, arXiv 2506.02153, first version June 2025 · Source
Notes
Latest engineering notes.
- A2A is not your internal message bus.
A2A is a protocol for agents run by different organizations to exchange tasks. When it is the right boundary, and when an internal queue is enough.
- Boring patterns that keep multi-agent systems up
What keeps multi-agent systems from collapsing into themselves: scoped context, explicit handoffs, failure budgets, observability, and a shutdown switch.
- Retrieval or fine-tuning: ask about the data first
When to retrieve and when to fine-tune, framed as four questions about your data — not your model. Plus the case for doing neither.
Equalize the gap.
Bring one decision. Leave knowing the first step.
A short description of the decision, who owns it, and where it runs is enough for the first conversation.
https://eqlx.ai/about · Printed copy; the website is the current version.