Retrieval layer
Your documents parsed, chunked and indexed with source references, so every answer can point back to the passage it used.
AI ASSISTANTS AND AGENTS
Assistants that do the work, not demos
Retrieval over your own documents, task automation with a human reviewing the output, and the evaluation set that tells you when it stops being accurate.
Test this against my documents
30 minutes with an engineer, no sales call
What this is
Answers grounded in documents you control
An assistant is useful when it reads your own material and shows where an answer came from. An agent is useful when it finishes a task and a person can see what it did. We scope both for accuracy and review: an evaluation set defines what a correct answer looks like, guardrails limit what tools the model can reach, and anything customer facing ships with a human in the loop.
This is for you if
- Your team answers the same questions from documents that live in four different places.
- Someone spends hours each week reading records, extracting fields and pasting them into another system.
- You tried a general model on your own material and the answers looked confident and were wrong.
This is not the right fit if the goal is a chatbot on the marketing site with no data behind it. Without your documents and a definition of a correct answer there is nothing to evaluate, and we would be shipping a widget that invents things about your company.
Scope
What we build
Tool access with permissions
The actions the model may take, scoped per user and per role, checked on the server before anything runs.
Evaluation set
Real questions with agreed correct answers, run on every change so accuracy is measured rather than assumed.
Review and escalation paths
A queue where a person approves, corrects or rejects output, and a defined route when the assistant should stop and ask.
Cost and latency controls
Model routing, caching and limits per request and per account, so spend and response time stay inside a bound you set.
Logging
Every prompt, retrieved source, tool call and decision recorded, so a bad answer can be traced instead of guessed at.
How it goes
From material to reviewed output
- 01
Task and accuracy definition
One week
We pick one task, gather the documents behind it and write down what a correct answer looks like with your team.
- 02
Retrieval and evaluation build
Two to three weeks
Indexing, source citation and the first evaluation set running on real questions, with results you can read.
- 03
Tools, review and guardrails
Two to four weeks
Permissioned tool access, the review queue, escalation rules, cost and latency limits and full logging.
- 04
Supervised rollout
Two to four weeks
A small group uses it with every output reviewed, and the evaluation set grows from the corrections they make.
Artefacts
What you receive
Evidence
Evaluation set in your repository
Evaluation results per change
Known failure cases, written down
Guardrail and escalation rules
The system
Source code in your own repository
Retrieval pipeline and re-index procedure
Review queue for human approval
Prompt and tool definitions under version control
Operations
Request and cost logging dashboard
Latency and spend limits per account
Runbook for adding documents and tools
Model change procedure with re-evaluation
Your documents stay in storage you own. Model providers are named in writing before anything is sent, and access can be revoked by you at any time.
Related work
The rest of this line
Custom platforms
Web platforms and internal tools built around how your business actually works.
Integrations and data
APIs, event pipelines and one reliable version of your operational data.
Back up one level
Software development
Custom platforms, AI assistants and the integrations that hold them together.
Assistants that answer with your own data and agents that finish real tasks.
Questions
Questions we get about assistants and agents
Next step
Start with one task and one set of documents
Book a technical assessment
30 minutes with an engineer, no sales call