AlgoMaker · institutional
From idea to filled order, with proof at every step.
A skeptical lab for AI operating in high-stakes environments. Out-of-sample gates, cost measured per instrument, the lot size the broker actually accepts, and the gap between what the test promised and what the desk delivered.
It runs on your infrastructure, with your key. Arming real capital stays a human click.
The shift
An agent should not have to pretend to be a mouse.
A robot moving a cursor breaks the day a button moves — and leaves nothing behind you can audit. The alternative is the desk exposing its own operations as functions: the agent calls, the mandate bounds what it may call, and every decision is recorded with its reason. That is the difference between a track record you can explain and one you can only show.
Buy. Sell. Drag the stop. Read the chart. Every decision routed through attention, fatigue and screen time.
data.download()
mine.strategies()
retest.funnel()
portfolio.run()
deploy.paper()
exec.status()The stack
Three layers, one pipeline, nothing skipped.
Research
Candidate generation across instruments and timeframes, then a funnel designed to kill rather than to confirm. Out-of-sample gates, walk-forward periods and Monte Carlo resampling run before anything is promoted.
- Market data ingestion and universe maintenance by liquidity
- Execution cost modelled per instrument, not assumed
- Statistical arbitrage on cointegrated pairs as a separate track
Portfolio
Surviving strategies are organised into banks and composed into portfolios, with correlation examined across the book rather than strategy by strategy.
- Composition, ranking and retest of an existing book against new data
- What-if analysis on parameter and cost assumptions
Execution
Paper first, always. Live execution runs against connected venues with position state, fills and realised cost measured against what the backtest assumed.
- Perpetual and spot venues, with per-venue fee and slippage tables
- Live divergence readout: what the test predicted against what the desk is delivering, per strategy
- Stop and kill available at all times, independent of the agent
Fleet intelligence
The data that matters is not the history. It is what happened next.
Price history is a commodity — every desk has the same candle. What almost nobody holds is the pair: the genome the machine proposed, and what executing it actually cost, asset by asset. The fleet returns that to research, and the gate tightens because of it — a strategy that only survives on an optimistic execution assumption stops being approved.
Slippage, spread and realised cost come from live fills across the fleet, per instrument and per venue — not from a constant typed into a config file.
Only anonymised percentages leave a machine, and only once enough independent sources exist for a figure to be non-identifying. Positions, balances, strategies and credentials never travel.
Fleet-measured cost becomes a prior in the research pre-flight. A strategy that only survives on optimistic execution assumptions stops being approved.
The population is inspectable: strategy families, the indicators they share, sample size, and where live delivery diverges from what the test promised.
Fleet view — each swarm is a strategy family, each point an execution, each thread an indicator shared between families. Illustrative.
Controls
Autonomy is the easy part. Restraint is the product.
Anything can be automated. What matters on a desk is what the system refuses to do, and what it records while doing the rest.
Arming live capital and registering venue credentials are human actions on a screen. The agent prepares, monitors and shuts down — it does not switch itself on.
Without a defined mandate — capital, maximum tolerated drawdown, objective, horizon — the research pipeline refuses to run.
Campaigns are checked against measured constraints before they consume hours of compute, and report why a configuration would return nothing.
Stop and force-close paths are exposed as first-class operations, callable independently of whatever the agent is doing.
Detail stays on the operator's machine. Only derived, anonymised aggregates leave it — never positions, credentials or balances.
The fee and slippage table used in research is the same one the executor charges on close. Research cannot be cheaper than reality.
Integration
It plugs into the model you already pay for.
The platform does not sell you intelligence. It exposes the desk, and your agent brings the reasoning — over your own subscription, under your own governance.
Native Model Context Protocol server. The agent discovers the capability surface and calls it directly, with tool-level descriptions and guardrails.
Authenticated HTTP interface on the operator's machine for systems that are not agent-based.
A human view over the same engine: research runs, portfolios, live monitor and the divergence plane.
Runs on the operator's own infrastructure, including isolated VPS nodes. No custody of client funds at any point.
Who this is for
Desks that already believe in process, not prediction.
Two rooms, one stack. The desk that already runs like an institution, and the operator building the one that will.
Already institutional
Proprietary desks, family offices, systematic managers and technology teams inside financial institutions. You have the mandate and the capital; what you want is the research-to-execution path machine-operable and auditable, running on your own infrastructure.
Becoming one
Operators running their own book who intend to raise. What separates you from a fund is rarely the strategy — it is the record: a written mandate, a funnel that refuses, execution measured against what the test promised, and an audit trail an allocator can read. The stack produces that as a by-product of running.
Who this is not for
If you are looking for signals, a managed account, or a return target, this is the wrong door — and it is better to know now. We are a technology provider: we do not advise, do not recommend assets, and never custody client funds. Most candidates the funnel generates are supposed to die before they reach capital.
Contact
Come see the pipeline kill a strategy live.
The demonstration is the funnel running: candidates generated, candidates discarded, and the handful that survive out-of-sample with cost applied. It is a more honest first meeting than an equity curve.
Request a technical demonstration