Custom AI engineering
AI agents built around your business.
We design and build custom AI agents that understand your company's knowledge, automate repetitive workflows and integrate with the tools your team already uses.
- Custom build, not a licence
- Typically live in 4–6 weeks
- Runs on your data
The difference
Not another chatbot.
A generic assistant knows the internet. It does not know your suppliers, your part numbers, your approval limits or the way your team actually closes a case. We build the second thing.
Every project starts from one process and the information behind it. That is what makes the output usable on a Tuesday afternoon rather than impressive in a demo.
- Your data
- The agent answers from your documents, records and knowledge base — not from whatever a general model happens to remember.
- Your workflows
- It is built around one real process, with the same steps, exceptions and hand-offs your team already follows.
- Your tools
- Where it helps, the agent reads from and writes to the systems you run today through their existing APIs.
- Your business rules
- Approval thresholds, wording, escalation paths: the constraints are explicit, and the agent stays inside them.
Solutions
Six shapes an agent usually takes.
Most projects start as one of these and then bend to fit the process behind it. The category matters less than the workflow it is built for.
Knowledge agent
Gives your team instant answers from your documentation, processes and internal know-how — with a link to the source every time.
retrieval
Document agent
Reads incoming documents at volume: extracts the fields you care about, classifies them and routes them onward.
extraction
Workflow agent
Runs a repetitive multi-step process end to end, and hands the case to a person at the points where judgement is required.
orchestration
Support agent
Answers customer and employee questions using your own policies, product data and past tickets rather than the open web.
resolution
Data agent
Queries your business data, explains what changed and turns the answer into a report someone can act on.
analysis
Email agent
Triages the shared inbox: classifies each message, pulls out the details and drafts a reply for review.
triage
Featured agents
What a finished agent looks like.
Four demonstration builds, shown with the interface a team would actually use. The data is illustrative — the structure is how we build.
What should I check if conveyor 17 stops detecting packages?
Conveyor 17 uses a retro-reflective photo sensor (SICK WL12-2) at the infeed. Work through it in this order:
- 01Clean the sensor lens and the reflector — dust build-up is the most common cause on this line.
- 02Confirm the alignment LED on the sensor body is steady green when a package passes; a flickering LED means the reflector has shifted.
- 03Check the sensitivity potentiometer against the 40–60% band recorded in the commissioning sheet.
- 04Measure the signal at terminal X4.2 in the control cabinet — under 18 V DC points to the cable gland at the drive end.
Two work orders in the last 90 days closed on the reflector bracket loosening after cleaning. If the fault repeats within a shift, replace the bracket (part 7742-B) rather than realigning again.
sources
- Conveyor 17 — Operating manualSection 5.4, p. 610.94
- WL12-2 sensor datasheetAlignment & sensitivity0.88
- Work order #48213Closed 14 Mar0.71
Technicians ask in plain language and get the relevant procedure, the fault history for that machine and a link back to the exact page of the manual — instead of paging through a shared drive while a line is down.
connected to
- Manual archive (PDF)
- CMMS work orders
- Shift handover notes
- Documents indexed
- 4,120
- Median answer time
- 6s
- Answers with a cited source
- 100%
How it works
Four steps, and you see something working in the third.
A first agent usually goes live in four to six weeks. The order matters: we do not connect anything until we agree on which process is worth automating.
01
Discover
We sit with the people doing the work and map one repetitive process end to end — where the time goes, where the errors happen, and what a good outcome looks like.
Week 1
- Process map
- Success criteria
- Feasibility call
02
Connect
We connect the information the agent needs: documents, knowledge bases, records and the systems your team already works in.
Week 1–2
- Data sources
- Access & permissions
- Evaluation set
03
Build
We design and develop the agent around that specific workflow — the retrieval, the tools it may call, the rules it follows and the points where it asks a human.
Week 2–5
- Working agent
- Review interface
- Guardrails
04
Improve
Once it is live we watch real usage, measure what it saves and keep refining answers, coverage and edge cases.
Ongoing
- Usage traces
- Quality reviews
- Monthly report
Use cases
The same building blocks, pointed at different work.
Retrieval, extraction, reasoning and tool calls are not industry-specific. What changes is the vocabulary, the documents and the rules — which is exactly the part we build for you.
Manufacturing
Machine documentation and fault histories made searchable on the floor.
Customer support
First-line answers drafted from your own policies and past tickets.
Logistics
Shipping documents read, checked and matched against orders.
Professional services
Contracts and reports summarised with the clauses that matter surfaced.
Operations
Recurring checklists and hand-offs run without the copy-paste.
Internal knowledge
Onboarding questions answered from the handbook instead of a colleague.
Administration
Invoices, forms and requests classified and routed on arrival.
Not on the list? The question we start with is always the same: which task takes your team the most time for the least judgement.
Why us
A small engineering team, working close to the problem.
No account managers between you and the people building the thing.
Built around your workflow
We do not push your business into a generic AI product and call the leftovers a limitation. The process comes first; the agent is shaped to fit it.
Your information, your context
Agents work with the documents, records and vocabulary your company actually uses — including the parts that only exist in a shared drive or someone's head.
Designed for real work
We agree on what success means before we build: hours returned to a team, cases closed without escalation, documents that no longer need retyping.
Honest about the edges
Some tasks should not be automated, and some answers need a person. We say so, and we build the hand-off rather than hiding the uncertainty.
Connected to your tools
Where it helps, agents read from and write to the software you already run through its existing APIs. Where it does not, we leave it alone.
Yours to keep
You get the implementation, the prompts and the evaluation set. Hosting can stay in the EU, and nothing is locked behind a seat licence you cannot leave.
Start here
Have a process that could be automated?
Tell us what your team spends too much time doing. We will look at it and say honestly whether an AI agent helps — and if it does not, what would.
Contact
Tell us about the process.
Describe the work in your own words. You do not need to know what kind of agent it should be — that is the part we figure out.
- Reply
- Within one working day
- First call
- 30 minutes, no slide deck
- After that
- A written view on whether it is worth building
- Based in
- Miami, Florida
Prefer email? hello@onion-agent.com