Build History

How Tantoryn AI Was Built

A curated, chronological record of pre-launch milestones — from the project's earliest working prototype through the Website v2 launch. It is not a log of every task or audit; it is the shape of how the system was actually built.

Build History is retrospective: every entry below happened before Website v2 launched. For what happens from here on, see the News page.

View News →
Milestone

Tantoryn AI Website v2 Launches

A new multi-language public website launched, with dedicated Technology, Research, Status, News and Build History sections replacing the earlier prototype.

Research

Raising the Bar for Future Research

An expanded methodology for future ML robustness testing — behavior under shifted, noisy or missing data, and across market regimes — was adopted as standing research policy.

Runtime

Trade History, on Request

A Telegram trade-history reporting feature was deployed to production, letting users request automatic and on-demand summaries of observed virtual trading activity.

Research

How the Layer Combines Model Outputs, Decided Carefully

A leakage-safe, independently audited procedure was adopted for selecting how multiple model outputs are combined into a single directional ranking.

Audit

A Negative Result, and What Came Next

A full independent-evaluation phase closed after methodological audit, finding the evidence not yet sufficient for Production status — and a new, tighter, bounded research programme began immediately.

Research

Comparing Candidates, Not Picking Favorites

A structured, multi-phase model-comparison programme evaluated several candidate approaches side by side under frozen rules, closing without granting any of them Production status by default.

Research

One Methodology, Not Several

A canonical, leakage-audited dataset and training methodology was unified across model families, replacing several parallel, slightly inconsistent approaches.

Research

A Full Model Family, Trained and Accepted

A full family of model architectures was trained, validated and accepted end to end on the second reference market.

Audit

A Second Market, the Same Discipline

A full research/audit cycle closed on a second reference market, clearing it for a full model-training cycle under the same evaluation discipline as the first.

Product

Opening the Doors: Public Beta

Tantoryn AI opened its Telegram-facing public beta: the first audience-facing observation stage, with real-money order execution explicitly disabled.

Engineering

A New Home for the Runtime

Infrastructure was migrated to a new production server, completing with zero data loss and improved network stability.

Milestone

One Candidate, Start to Finish

The first fully automated proof that a live market candidate could move through generation, ML scoring, risk approval, and virtual trade open and close without manual intervention.

Runtime

The First Overnight Run

The system completed its first overnight live-paper run — roughly fifteen hours and hundreds of automated cycles — proving basic runtime resilience under real network conditions.

Research

Choosing What Stays In

After comparative testing, the production candidate-family set was finalized, and at least one experimental family was retired from production based on the evidence.

Milestone

Pipeline Connected, End to End

The Candidate Generator, ML evaluation and Risk/Lifecycle stages were connected into one pipeline and validated out-of-sample on a reference market for the first time.

Audit

The First Audit Gate

The first full data-quality audit gate closed on a reference market dataset, correcting multi-timeframe merge logic and label definitions before further ML work proceeded.

Engineering

A Backtest Engine Honest About Costs

A paper-trading runtime and a realistic backtest engine were built, modeling transaction costs, slippage and execution delay instead of assuming frictionless fills.

Engineering

Rebuilding the Market-Data Foundation

The market-data pipeline was rebuilt and expanded to include liquidity, macro and on-chain sources, and the first baseline ML training cycle ran on a reference market.

Research

First ML Models, and a Lesson About Labels

The first machine-learning model stack was assembled — and an early flaw in how training labels were defined was found and corrected before it could mislead later work.

Milestone

The First Line of Code

Tantoryn AI's earliest working system: a basic data pipeline and trading-bot framework covering data collection, risk management, asset filtering and reporting.