Start trading, not building

What you would have to build

A mid-frequency platform is eight systems that have to agree with each other. Each row is a part Nexus runs today, with the page that shows it.

• Market dataCapture · Reference data

  Ten sources and nine asset classes, about 1.28 billion bars from daily history back to 1969 to live one-second capture. Futures contracts rolled and back-adjusted by a recorded rule, with exchange calendars, roll rules and instrument records kept as dated ranges.

• Data qualityValidate and clean

  Raw data kept as delivered, 37 check types, every fix a logged and reversible operation, and bad series quarantined until fixed.

• Research under controlsResearch · Tester

  Six simulation engines and a strategy tester, with trials counted before any data is read, sealed hold-outs, costs charged on every trade, and an authorisation no agent can write.

• Model interface and sizingModels

  One contract for every model, and one sizing library shared by the backtest and the live book, so the two cannot disagree about how a forecast becomes a position.

• Execution and riskExecute

  A per-market scheduler, staged pre-trade limits that never block a risk-reducing order (production runs them at the warn stage), a halt and a kill switch, reconciliation with the broker, cost analysis, and one interface to the broker.

• OperationsProduction

  43 checkers and seven fast checks, failure alerts wired on 321 of 355 services and timers, red checks that become work items, and backups that are restored on a schedule.

• DeliveryCI · Agents

  Tests chosen by what a change touches, self-hosted runners, merge lanes, independent review of the exact commit, and a deploy that refuses a dirty tree or a rollback and logs every release.

• Consolefor the desk

  A web console of about 80 pages over all of it: data, research, the book, orders, controls and health.

The team it takes, and what it costs

An estimate for a firm that hires its own team in London to build the same breadth. It is a scale, not a quote: your mix of roles, seniority and location will move it.

• £1.4m–£3.7m

  a year to employ the team, adding about 30% to pay for employer National Insurance, pension and benefits.

• 3–6 months

  to hire. Notice periods and gardening leave set the pace for people who have done this before.

• 12–18 months

  to build to this breadth with these controls. Practitioners quote 9 to 18 months for a two-person team on a much narrower setup, and first estimates tend to slip.

• £1.6m–£6.3m

  to a first demo trade: 12 months of the lean team's employment cost at the low end, 18 months of the full team's at the high end, plus recruitment fees of 20–25% of first-year pay.

• £0.9m–£2.5m

  a year afterwards, for three to five people to run and extend it: a lead, a quant developer and a platform engineer, up to the lean team of five.

Not counted: data licences, servers, office space, the months spent hiring, and the founders' own time. Not counted either: the model earns nothing while the platform is built around it.

Why not estimate from the code? Nexus has about 510,000 lines of production code and 320,000 lines of tests in its core repositories, and code-size models would put the effort far higher than the team above. Most of that code was written by AI agents, which write more lines than people do for the same function, so a count of lines overstates the work. The estimate above starts from the team instead.

Built to stay cheap to change

A platform's real cost is what it takes to support and change it after the first release. Nexus was designed and built from the ground up with the practices that keep that cost down. Each one below is evidenced on a page of its own.

• Clear interfacesModels · Broker

  Models come in through one contract and orders go out through one broker interface. Research and trading read prices through one data API and size positions with one library. Major decisions are written down as design records.

• Tested by defaultCI

  27,275 test functions. Every change runs the tests it affects, and anything the maps do not cover runs the full suite. Every deploy passes its service's test gate first. Mutation tests check that key safety tests can actually fail.

• One way to shipAgents

  Every change is a work item, made in its own worktree and approved by a different reviewer at the exact commit, then merged through a lane and deployed by a tool that refuses a dirty tree or a rollback. Fixes go forward.

• Audit at every layerrecords you can query

  A work board with 8,025 recorded leases and a deploy ledger with 1,705 releases (both to 15 October), an append-only register of research trials, 6,631 logged and reversible data fixes (to 1 October), a record of every order, and model versions identified by the hash of their declaration.

• WatchedProduction

  Checkers turn anything red into a work item, failure alerts are wired across the services and timers, and each production incident has left a rule or a tool behind.

Reviewed three times against prop-desk due-diligence standards, the platform's mean score rose from 2.2 to 3.4 out of 5, with 4 for code quality, testing and CI, and 4 for change management and rollback. The reviews are self-assessed and say so; the method and all 13 scores are published.

Ready today, and what a first engagement adds

Nexus is a running platform, not a kit of parts. A new firm still has a short list of steps of its own, and the list is part of the offer.

Who built it

A trading systems engineer with more than 20 years in financial services. I joined Marshall Wace in 2002 as the CTO's first hire, a founding member of the team that built its early systematic trading services, and have since worked on the desk across the whole life of a systematic strategy: onboarding models at Marshall Wace and Ashenden, building a portfolio management system from the ground up, and running trading operations for a high-frequency crypto market-making pod at Tower Research Capital. Nexus applies what those years taught about which systems stay supportable, and which become too expensive to change. More about me.

Stated plainly

• Demo only. Execution runs on broker demo accounts. Nexus has not traded real money, and no live account has placed an order.
• One operator. One person directs the AI agents that build and run it. The bus factor is one.
• AI writes most of the code. The architecture, controls, review standards and rulings are mine.
• Coverage is measured in one place so far. The trading service's build fails if its tests cover less than 68% of its code (82% on 28 October 2026), or if any of its 14 risk-control and calculation modules drops below its own floor. Elsewhere the tests are counted, and the market-data suite fails if its count drops below a floor, but coverage isn't measured yet.
• The estimate is an estimate. It rests on published pay data and the build times above, not on a quote for your firm.
• The scores are self-assessed. A 5 needs someone who did not build it.