01
Discover
What you buy A process map, a ranked pain-point register, a quantified baseline and an implementation proposal.
Evidence to continue A named sponsor, a useful problem, viable access and a budget.
00 For Forex and CFD brokers
Private AI, automation and integrations, delivered on site by a forward-deployed engineer. We find the workflow that actually costs you time, build the smallest thing that improves it, and show the evidence before the scope grows.
Fixed-fee discovery
A written assessment you can use even if you do not proceed to implementation.
Broker-controlled data
A verified data boundary, with every exception written into the contract.
Measured before scaled
A baseline, representative cases, and a decision — not a platform sold in advance.
01 Where the time goes
These are the recurring operational loads we look for in a brokerage. The first engagement starts with whichever one your own staff can show us, with a recent example and a sequence of handoffs.
01
Staff search policies, tickets and colleagues for an answer that should already exist. Entity rules conflict with last month’s version.
Typical buyer · COO or head of support
02
Onboarding and periodic refresh stall on incomplete files, duplicate entry, expired documents and unclear ownership.
Typical buyer · Compliance or operations
03
CRM, client portal and trading data are copied by hand. Identifiers drift. Failed handoffs show up late, usually in a spreadsheet.
Typical buyer · CTO or COO
04
Client and policy data cannot leave an agreed environment. A general AI tool does not meet that requirement, and a local model alone does not either.
Typical buyer · CIO and security
02 How we work
Each stage is a purchase with a deliverable and a reason to continue. We do not ask a broker to buy a platform before one workflow has proved itself.
01
What you buy A process map, a ranked pain-point register, a quantified baseline and an implementation proposal.
Evidence to continue A named sponsor, a useful problem, viable access and a budget.
02
What you buy One integrated prototype in a controlled environment. A conventional automation is equally valid when a language model adds little.
Evidence to continue Measured quality and improvement on representative cases.
03
What you buy A hardened application, staff training, runbooks and a handover your own team can operate.
Evidence to continue Signed acceptance, an operating owner and support arrangements.
04
What you buy Defined support, updates and a bounded improvement backlog. A recurring fee buys a continuing service, not an indefinite subscription for a one-time install.
Evidence to continue Actual usage, reliable operation and a justified renewal.
05
What you buy Adaptation of a proven module for another broker or another entity. Customer records and policies are never reused without authorisation.
Evidence to continue Similar demand, and an implementation that is materially reusable.
03 The discovery week
This is an embedded workflow assessment, not an informal visit. We need people, examples and system documentation. Unrestricted production access is not required. Preparation and the written report sit outside the onsite days.
Day 01
Meet the sponsor and the process owners. Establish the operating environment.
Leaves with System inventory, workflow boundaries, success criteria.
Day 02
Observe staff handling ordinary cases and the exceptions that actually consume the day.
Leaves with Current process map, handoffs, an initial time baseline.
Day 03
Review documents, integrations, permissions and hosting. Start from synthetic or redacted examples.
Leaves with Data-flow map, API feasibility, security constraints.
Day 04
Compare candidate improvements with the people who do the work and the people who own the systems.
Leaves with Ranked opportunities and a small proposed pilot.
Day 05
Review findings, assumptions, cost and who is responsible for what.
Leaves with Discovery report and an agreed next-step decision.
04 What we build
We integrate the systems you already run. Replacing the CRM, the trading infrastructure or the payment stack is not the opening offer. The model is never the authority that grants a permission or approves a sensitive transaction.
01 · Launch
Problem Staff cannot find the current policy for the right entity.
Build A staff assistant that searches approved documents, shows the source, and prepares drafts for review.
Control Permissions and entity boundaries apply before retrieval. Missing evidence is escalated, not invented.
02 · Launch
Problem Onboarding and refresh create rework: incomplete files, unclear owners, expired documents.
Build Intake, extraction with confidence, completeness checks, expiry alerts, review queues and version history.
Control Originals stay original. Broker staff keep the acceptance decision. The model does not repair a date of birth or manufacture evidence.
03 · Launch
Problem People copy between CRM, portal, documents and trading systems.
Build A small adapter set: field mapping, scheduled or event-driven jobs, reconciliation and a visible error queue. MCP is used where it simplifies approved tool access. A direct API is used where it does not.
Control A successful read does not silently become permission to change trading or account records.
04 · Foundation
Problem The workflow needs a model, and the records must stay inside an agreed boundary.
Build A broker-controlled runtime with retrieval, identity, monitoring and one application. Candidate models are benchmarked before hardware is bought.
Control Inference, OCR, embeddings, logs, backups and support sessions are all in the data-flow review.
05 · Expansion
Problem Spreadsheets disagree, and decisions wait on the argument.
Build Dashboards for onboarding, support, document refresh and backlogs. Natural-language questions only over a governed reporting layer.
Control Queries stay inside authorised data. Definitions and freshness are published. Totals reconcile to the source systems.
06 · Selective
Problem A portal, migration or connector is genuinely needed, and an existing product will not do it economically.
Build A tight scope, a working prototype, user validation and milestones. Customer-specific logic stays out of the shared core.
Control Supported systems, data ownership, acceptance and an exit are agreed before the work starts.
05 Private AI, defined
Retrieval-augmented generation pulls approved source material into the answer. A system prompt sets behaviour. Fine-tuning changes model weights. They are different tools. We start with an existing model, prompts and retrieval, and commission fine-tuning only when a measured shortcoming justifies the cost.
Authenticate staff. Connect approved records. Enforce entity, role and document permissions.
Parse, index and retrieve approved documents. Keep sources, versions and access restrictions.
Run the selected model inside the agreed boundary. Benchmark quality, languages, latency and cost.
Search, draft preparation and workflow tasks, with clear actions, review queues and evidence.
Connect CRM, documents and other approved systems. Narrow permissions. Validate every write.
Monitor, update, back up and recover. Named owners, change control and a tested restoration.
On premises
The runtime sits in the broker’s own environment. Patching, capacity, licences and recovery are part of the design, not an afterthought.
Private cloud
A dedicated environment the broker controls. A private cloud is not a server in the office, and we do not describe it as one.
Managed region
We operate the environment in an approved region. Hardware, inference, monitoring and support are priced as themselves.
What “data stays local” must mean
We define exactly which data stays where. The review covers inference, OCR, embeddings, logs, telemetry, backups, support sessions, external screening providers and update mechanisms. Unapproved outbound traffic is disabled and checked. If a KYC provider receives documents, or a backup leaves the premises, the contract says so. An absolute claim that nothing ever leaves would be false, and we will not make it.
06 After a broker is qualified
Procurement, hardware delivery and access approvals can extend this sequence. A prototype shows feasibility. A pilot tests value with a limited group. Production adds the operating controls. A successful API call is none of those three.
Weeks 1–2
Scope, access and a baseline. Participants are named before anyone arrives.
Week 3
The forward-deployed engineer works with process owners. The week ends with an agreed problem and a candidate workflow.
Week 4
APIs and the deployment boundary are checked. The sponsor approves the economics and the acceptance criteria.
Weeks 5–7
Built in a controlled environment. One entity, one team, a defined dataset, a small set of permitted actions.
Weeks 8–9
Representative work runs beside the existing process. Quality and benefit are written down, including what did not improve.
Weeks 10–11
Training, recovery checks and a production-readiness decision by the technical and operational owners.
Week 12
Acceptance and handover, with a support scope that matches what was actually built.
A first pilot we prefer, once discovery confirms it: an internal policy and support assistant, or document-expiry management with a staff review queue. An illustrative acceptance starting point for a knowledge pilot is a 20% reduction in median handling time without a drop in reviewer-rated correctness. That is a negotiating position, not a result we claim in advance.
07 Indicative fees
Figures are planning ranges in US dollars, excluding tax, travel beyond the stated allowance, and hardware or cloud the broker owns. The contract fixes the number. A pilot is priced so its cost is covered even if it does not convert.
Discovery
from $6,000
Focused pilot
from $20,000
Managed improvement
from $2,000/month
Broader scope, unfamiliar APIs or additional entities are estimated separately. Deposits are collected before we make material supplier commitments. Fees are not tied to trader losses or to trading performance.
08 Scope
The opening pilot stays inside an internal workflow, with controlled access and actions a person can review. Anything below needs its own case, its own operating design and the applicable approvals. It is not added quietly to an implementation contract.
09 Who does the work
The first discovery and the first implementation are led by the founder, with contracted specialists where the job needs them. Buyers can already reach AI tools, and many have capable engineers. The advantage we offer is operating knowledge of a brokerage, connecting systems safely, taking delivery responsibility, and writing down the result. That advantage is earned in the engagement. It is not established by a label.
From the first deployment we separate a reusable core — identity, document ingestion, permission-aware retrieval, evaluation, audit events, deployment — from the adapter that belongs to one broker. A second customer should not mean rebuilding everything. It also should not mean reusing another broker’s records.
10 The first step
The first step is a fixed-fee discovery week. You leave with a process map, a baseline and a costed proposal — useful whether or not the pilot goes ahead.