Offer 04 · Custom AI systems
AI systems engineered for the problems no template fits
Vision, forecasting, bespoke models and enterprise platforms, built to the edge cases of your business and owned by you.
- Partial invoice with a currency changepassed
- Order placed during a stock recountpassed
- Duplicate sensor reading after a restartpassed
A case that fails stops the release until it is fixed or approved.
Sample data. Representative workflow, not a client screenshot.
Types we build
Six kinds of system, each built for one decision
- 01
Computer vision inspection
Checks images or video against your standard, and sends the uncertain ones to a person.
Build this agent- Image arrives
- Model scores each check
- Uncertain case goes to review
Example workflows
- A line camera that flags misaligned seals and logs each reading.
- A document check that compares a scanned form against the expected layout.
- 02
Demand forecasting
A forecast with a range, so planning shows how much to trust each number.
Build this agent- Sales and stock data
- Forecast with a range
- Planner adjusts and approves
Example workflows
- A weekly stock forecast by site that widens the range when demand is volatile.
- A staffing forecast that flags the weeks where the model has little history.
- 03
Bespoke scoring model
A model trained on your own history for a decision no vendor product covers.
Build this agent- Case arrives
- Model scores with reasons
- Borderline case goes to a person
Example workflows
- A credit-review score that lists the factors behind each result.
- A lead-quality score trained on your past wins and losses.
- 04
Enterprise AI platform
Shared model services that several teams use, with access and audit built in.
Build this agent- Team requests a model
- Platform checks access
- Call and result are logged
Example workflows
- An internal model service that several departments call with their own permissions.
- A shared evaluation service that tests each team's model before it goes live.
- 05
Pricing model
Suggests prices from your history, with each factor recorded in the audit trail.
Build this agent- Quote requested
- Model scores with factors
- Deviation approved by a person
Example workflows
- A quote tool that suggests a price band and lists the factors behind it.
- A discount check that routes unusual offers to a manager and logs the decision.
- 06
Anomaly detection
Flags unusual transactions or readings, and logs each flag with its evidence.
Build this agent- Reading arrives
- Model compares to normal
- Alert and audit record
Example workflows
- A transaction monitor that flags unusual payments and keeps the evidence for review.
- A sensor watch that records each alert with the readings that triggered it.
Where templates fall short
Four places a general model fails your business
01
The off-the-shelf model misses your edge cases
- Agent does
- The model is trained and tested on your own edge cases, and each release shows the cases it gets wrong.
- Person approves
- Any release that changes the decision threshold.
BeforeAfterA general model is trained on common cases. Your unusual ones fail quietly, and the failures look like normal results.
- Vendor model scored an unusual order as normal
- Exceptions were found by customers, not by the system
- Nobody could say how often it was wrong
Edge cases, tested before release- Partial invoice with a currency changepassed
- Order placed during a stock recountpassed
- Duplicate sensor reading after a restartpassed
A case that fails stops the release until it is fixed or approved.
Sample data. Representative workflow, not a client screenshot.
02
The forecast looks right until something changes
- Agent does
- Each forecast carries a range, and the range widens where the history is thin or the signals are unstable.
- Person approves
- The planning range the business acts on.
BeforeAfterA single-number forecast hides the risk. It is correct on an ordinary week and wrong the week a supplier fails.
- One number presented as the plan
- Forecast missed a supplier delay by a month
- No one saw the range until after the stock-out
Demand forecast, next 8 weeksThe range widens with each week, so the plan shows how much to trust each number.
Sample data. Representative workflow, not a client screenshot.
03
A vision check that nobody can audit
- Agent does
- Each detection is stored with its confidence and the image it came from, and uncertain results go to a person.
- Person approves
- Changes to the acceptance standard on the line.
BeforeAfterA camera flags defects, but nobody can say why one part passed and the next was rejected.
- Rejects with no recorded reason
- Operators overrode the system silently
- Audit asked for evidence that did not exist
Visual inspection, line 3- Scratch, left edge0.97Logged
- Seal misaligned0.71Sent to a person
- Surface clean0.99Passed
Sample data. Representative workflow, not a client screenshot.
04
The model runs on a vendor account you cannot move
- Agent does
- The model, training data, inference and logs run in accounts and storage your business controls.
- Person approves
- Where the system runs and who holds access.
BeforeAfterThe system works, but the model, the data and the logs sit with a vendor, and leaving would mean starting again.
- Model only runs in the vendor's account
- Training data uploaded and never returned
- Cost to leave is higher than the cost to stay
Where it runs- Model weightsYour account
- Training dataYour storage
- InferenceYour infrastructure
- Logs and auditYour retention rules
Sample data. Representative workflow, not a client screenshot.
Questions buyers ask
Before you scope a project
- What is a custom AI system?
- It is an AI model or service built for one decision, trained and tested on your own data and edge cases. It is engineered for your problem rather than adapted from a general product, and you own what gets built.
- When should we build rather than buy a vendor product?
- When the vendor product misses a case that matters to your business, or when the data cannot leave your control. We test the vendor option against your edge cases during scoping, so the decision rests on evidence.
- How do you test an unusual case?
- We write the edge cases down with your team and test the model against them before each release. A failing case stops the release until it is fixed or a person approves the risk.
- Who owns the model?
- You do. The model, training data, code and deployment are handed over to you. Nothing is held in a vendor account as a condition of using the system.
- How is the price set?
- Each phase has a fixed price and a written scope that you approve before the phase starts. The first phase covers the data review and the evaluation plan.
- What happens after launch?
- Monitoring and retraining are a separate phase, scoped and approved the same way as the build.
- What does the audit trail record?
- Every action the system takes: the input, the source or factors used, the output, and who approved it if a person was involved. The trail is kept in storage you control, so your team can answer an audit question from the record.