Meetrum is Navtechy's research direction toward a self-evolving AI foundation model for autonomous trading. It is an ambitious problem space, and we want to be clear-eyed about it: this article explains the concept, why the problem is genuinely hard, and the principles that guide how we approach it — not a promise of returns, because anyone promising those in trading should be treated with suspicion.
What "self-evolving foundation model" means
A foundation model is a large, general model trained broadly enough to be adapted to many tasks rather than built for a single narrow one. "Self-evolving" is the harder idea layered on top: a system designed to keep learning and adapting as its environment changes, rather than being frozen at training time and slowly going stale. In a domain as non-stationary as markets — where the patterns that worked last year can stop working this year — the ability to adapt is not a nice-to-have; it is the entire challenge. A model that cannot evolve is a model with an expiry date.
Why autonomous trading is so hard
Markets are one of the most difficult environments to build AI for, for reasons that are worth being honest about:
- The ground keeps moving. Financial markets are non-stationary. A model that fits yesterday's behaviour can be actively wrong tomorrow, because other participants adapt too — you are not predicting a fixed system, you are competing inside an adaptive one.
- Signal is faint and noise is enormous. Most of what looks like a pattern is randomness, and overfitting to it is the default failure mode. The easier it is to find a pattern in historical data, the more suspicious you should be of it.
- Feedback is slow, noisy, and expensive. You cannot cheaply label "good decisions"; the market tells you slowly, and sometimes a good decision looks bad in the short run and vice versa.
- The stakes are real. A confidently wrong system does not just produce a bad answer — it loses money. Risk management is not a feature bolted on at the end; it is the foundation everything else sits on.
The principles guiding Meetrum
Rather than chase a magic model, our approach is built on principles that apply to any serious AI system operating in a hostile, changing environment:
- Adaptation over memorisation. The goal is a system that updates its understanding as conditions change, not one that memorises historical patterns and assumes they will repeat. Memorisation is exactly what fails when the regime shifts.
- Rigorous, adversarial evaluation. Every idea is tested against out-of-sample data and realistic conditions — costs, slippage, and the awkward periods, not just the flattering ones — because in this domain it is dangerously easy to fool yourself with a clean backtest.
- Risk-first design. Position sizing, drawdown control, and knowing when to do nothing are treated as core capabilities. The ability to stay out of a market you do not understand is as valuable as the ability to act in one you do.
- Humility about uncertainty. A system that knows when it does not know — and steps back — is worth more than one that is always confident. Calibrated uncertainty is a feature, not a weakness.
Why a foundation-model approach
Traditional trading models are often narrow and brittle: built for one instrument or one regime, and quietly obsolete when conditions change. A foundation-model approach aims for something more general — a base of broad capability that can be adapted, combined with mechanisms to keep adapting. That is a harder thing to build, and we treat it as a genuine research problem rather than a solved one, which is precisely why the discipline around evaluation and risk matters so much.
Where this fits in Navtechy's work
Meetrum is the most research-heavy end of what we do, but it draws on the same foundations as our product work — RAG, agentic systems, evaluation discipline, and shipping real software that runs in production, like Restrofi, InvoiceAI, and LeadsBuck. Ambitious R&D and pragmatic product building are not separate tracks for us; each sharpens the other. The habits that make a support bot trustworthy — grounding, measurement, honesty about uncertainty — are the same habits a trading system lives or dies by, only with far higher stakes.
An honest note
We are deliberately not publishing performance claims here. Autonomous trading is a field where confident promises are a red flag, and the responsible way to talk about it is in terms of approach, discipline, and principles — which is exactly what this is. Meetrum represents where our curiosity and capability are pointed, built on the same rigour we bring to everything.
If you are working on hard AI problems in changing, high-stakes environments and want a team that treats evaluation and risk as seriously as capability, that is the kind of work we find genuinely interesting.
