VegaForge

AI trading / Financial markets

AI trading systems,
built as infrastructure.

We build the harness that turns a model into a trading system: data pipelines, calibration, execution, risk limits and the infrastructure in between. For trading teams, funds and market operators.

See the system

System map

One model / five layers

01DataVenues, reference data, corporate actions, snapshots.
02ResearchSignal, calibration, backtest, a baseline to beat.
03RiskLimits, exposure, drawdown, a kill switch.
04ExecutionVenues, order types, impact, reconciliation.
05MonitoringTraces, P&L attribution, drift, alerts.
ModelOne input, not the system

The failure

The model is the smallest part.

Most teams start with a model and an API key. That is the first week of a project, not the product. The market does not care about your backtest.

A signal with no cost model

Fees, spread, market impact and borrow

A cost-aware edge

A backtest with leakage

Time-ordered folds, survivorship, regime shifts

A test you can replay

A model with no desk

Execution, risk, reconciliation, monitoring

The desk around it

The system

One model. Five layers. One owner.

The model is one box. The rest is the system that decides whether it is allowed to act.

Data

01
  • Venues
  • Reference data
  • Corporate actions
  • Snapshots

Research

02
  • Signal
  • Calibration
  • Backtest
  • Baseline

Risk

03
  • Limits
  • Exposure
  • Drawdown
  • Kill switch

Execution

04
  • Venues
  • Order types
  • Impact
  • Reconciliation

Monitoring

05

Traces, P&L attribution, drift, alerts, on-call.

Reconciliation

06

Ledger, venue and chain compared after every order.

Model

07

A signal or a probability. One input, not the system.

The work

Three ways to start.

You do not have to start with live capital. Most teams should not.

01

Research scanner

Find the gap before risking capital. A ranked list of opportunities with the evidence and the costs attached.

What you get

  • Universe scan
  • Signal with sources
  • Cost and liquidity model
  • Ranked shortlist

What it does not do

It places no orders and takes no exposure.

02

Execution agent

The model proposes. The harness disposes. Orders inside explicit limits, paper first, human approval where the size justifies it.

What you get

  • Sizing and limits
  • Paper trading
  • Human approval
  • Kill switch

What it does not do

It does not promise returns, and it does not hide the downside.

03

Risk and monitoring

Exposure, drawdown, reconciliation errors, failed orders and fee drag in one place. A short daily brief instead of another dashboard.

What you get

  • Exposure and drawdown
  • Reconciliation
  • Fee drag
  • Daily brief

What it does not do

It is not a dashboard nobody opens. If it does not change a decision, we cut it.

Before live capital

Six gates.

If one of them is missing, the model does not trade. None of them is optional.

01

Data you can replay

Snapshots, timestamps and the version of the rules at the moment of the decision.

02

A test set you have not touched

Time-ordered folds, a locked period, no peeking at the answer.

03

A cost model, not a mid-price fantasy

Fees, spread, impact, borrow and failed transactions.

04

A risk limit for every position

Size, daily loss, venue, correlation. Written down, not implied.

05

A kill switch that works

Tested, not assumed. One person can stop the system.

06

An owner for reconciliation

Ledger, venue and chain, compared and signed off.

The risk register

What we look for before it looks for us.

Lookahead leakageTime-ordered folds and a locked test set.
SurvivorshipDelisted and voided markets stay in the sample.
OverfittingPre-registered variants and a correction for the number of attempts.
Fee blindnessA cost-aware sizer, not a mid-price fill.
Market impactLiquidity and participation limits, sized before the order.
Broken feedStale-data guards, fallbacks and a visible failure state.
Model driftCalibration monitoring, retraining rules and a rollback path.
Reconciliation driftThree-way checks and one owner for the difference.

From production

We have run this harness on prediction markets.

The same layers apply to real markets. No client names, no P&L promises.

See the event market work

A weather ensemble priced against a market.

An LLM forecaster with a calibration layer.

A fee-aware sizer that survived live.

Read the harness article

How we start

Diagnostic, prototype, live.

No live capital on day one. The first deliverable is a decision, not a trade.

Week 0 to 2

Diagnostic

We audit the data, the signal, the execution path and the risk limits. You get a short report: what is missing, what to build first, what not to build.

Week 2 to 6

Prototype

A scanner or a paper-trading agent on your data. No live capital until the evaluation is clear and the failure modes are named.

Week 6 and later

Production

Live system, monitoring, reconciliation and a retainer. We stay on the failure modes, not just the launch.

Reading

All writing ↗

Your Polymarket Bot Is a Demo Until It Has a Harness

What the harness is, and what breaks without it.

The Five Things Polymarket's Winners Do and 87% Don't

What 112,000 wallets say about sizing, focus and exits.

The Weather Bots Quietly Making Millions on Polymarket

Why the boring markets attracted the best data pipelines.

The harness is the moat

Why the model is a commodity and the system is the advantage.

Start with a diagnostic

Bring us the market. We will build the harness.

Tell us what you are trying to trade and what you have already tried. We will answer with a point of view, not a sales deck.

alex@vegaforge.dev ↗

VegaForge / AI trading