AI agents
AI agents that answer from your own data
Veraplot builds AI agents and RAG assistants that answer from your own documents, databases and lakehouse, running on Amazon Bedrock inside your AWS account so your data stays within your perimeter. We combine Claude and other LLMs with vector search, tool use through MCP servers, text-to-SQL and guardrails, and we measure answer quality before anyone relies on it. Each agent is delivered at a fixed price by senior engineers, with the cost per conversation visible from the pilot onwards.
Grounded answers
Every response is built from retrieved passages or query results, with citations users can open. No source, no answer.
Your data stays yours
Models are called through Amazon Bedrock in your account and region. Your prompts and documents are not used to train the models.
Measured, not guessed
An evaluation set built with your experts scores accuracy, refusals and cost per answer, and is rerun on every release.
What you get
Use-case framing and evaluation set
The questions the agent must handle, the sources it may use, and reference answers agreed with your domain experts.
Retrieval pipeline
Document parsing, chunking, embeddings and a vector index in OpenSearch or pgvector, refreshed automatically when sources change.
An agent with tools
An agent that searches documents, runs read-only SQL on your lakehouse and calls your APIs through MCP servers, each with scoped permissions.
Guardrails and access control
Bedrock Guardrails, PII filtering and document-level permissions, so each user only retrieves what they are already allowed to read.
Interface, tracing and cost dashboard
A chat interface, a Teams or Slack integration, or an API, plus tracing, token usage and cost per conversation in one view.
The stack
Picked for your context, not for a partner badge. Open formats first, managed AWS services where they pay for themselves.
The brain
The right model per task, swappable without a rewrite
- Amazon Bedrock
- Claude
- Amazon Nova
- Mistral
- Llama
The memory
Retrieval over documents and the lakehouse
- Amazon OpenSearch Service
- PostgreSQL + pgvector
- Amazon Bedrock Knowledge Bases
- Amazon Titan Embeddings
- Amazon Textract
The hands
Tools the agent can call, with scoped rights
- Model Context Protocol (MCP)
- Amazon Bedrock AgentCore
- Strands Agents
- LangGraph
- Amazon Athena
- AWS Lambda
The guardrails
Scope, privacy and identity enforced in code
- Amazon Bedrock Guardrails
- Amazon Cognito
- AWS IAM
- Amazon Comprehend
- AWS KMS
The scorecard
Quality and cost, tracked release after release
- Ragas
- LLM-as-a-judge
- Langfuse
- Amazon CloudWatch
- AWS Cost Explorer
Typical use cases
Search and Q&A over a document platform
Industrial group. Natural-language questions over documents already indexed in the data lake, answered with links to the exact source files and respecting existing permissions.
Text-to-SQL on the Gold layer
Business users ask questions in plain language; the agent writes read-only Athena SQL against documented Gold tables and shows the query it ran.
Back-office copilot
An agent reads incoming emails or files, extracts the key fields, checks them against the ERP through an API and drafts the next action for human approval.
Internal knowledge for support teams
Procedures, contracts and past tickets searchable in one place, with answers limited to what each team is entitled to see.
From first call to production
- 01
Discovery call
We map your systems, constraints and the outcome that matters. You talk to the engineer who will build it, not an account manager.
- 02
Fixed-price scope
A written scope with deliverables, milestones and one price, agreed before any code is written. No day-rate meter running in the background.
- 03
Build in short iterations
Working increments every one to two weeks, demoed on your data in your AWS account. Course corrections happen early, not at the end.
- 04
Handover and optional run
Code, infrastructure as code, runbooks and a walkthrough with your team. Run it yourselves, or keep us on support month-to-month or on a 1–3 year plan.
Questions buyers ask
How much does an AI agent cost to build and to run?
The build is a fixed price, scoped once the sources and target questions are known. Running costs are separate and paid directly to AWS: Bedrock tokens, the vector store and light compute. We measure the cost per conversation during the pilot and design for it, with model routing, caching and tuned retrieval, so you pay for answers rather than wasted tokens.
Does our data leave our environment or train someone else's model?
No. The agent runs in your AWS account and region and calls models through Amazon Bedrock, which does not use your prompts or outputs to train models. Documents, vectors and logs stay in your own storage, encrypted with your keys, and access follows your existing identity provider. Nothing is routed through a third-party chatbot service.
How do you stop it from hallucinating?
By architecture rather than by prompt. The agent answers only from retrieved passages or query results and cites them; when retrieval finds nothing relevant, it says so instead of improvising. Guardrails block out-of-scope topics, and an evaluation set built with your experts is rerun on every change, so regressions are caught before users ever see them.
RAG, fine-tuning or text-to-SQL: which approach do we need?
RAG for documents and knowledge that change often; text-to-SQL for figures that already live in tables; fine-tuning rarely, and only for a specific style or output format. The most useful agents combine the first two: they search documents for context and query the lakehouse for numbers. We decide per question type during the framing phase.
How quickly can we have something usable?
A pilot on one well-defined use case typically reaches real users in four to ten weeks, covering framing, the evaluation set, the retrieval pipeline, the agent and its interface. The main variable is how clean and accessible your sources are, which is why many teams pair this work with a data platform engagement.