DRYCODEWORKS

AI systems: AI integration services

AI that could leak company data, give wrong answers or run up the bill: most of the work is making it safe for customers and staff to use. I build AI assistants that work on your company's data, AI features inside your product, and models trained on your own data, with permissions, spending limits and checks on what the model says.

Book a callSee client work

What I build for

Security and permissions

Permissions belong outside the model, where no prompt can change them, and in a multi-tenant product that includes keeping one customer's data out of another's answers. I fit the approach to the system: each user's own sign-in carried through to the data, permissions based on who owns what, row-level rules in the database, or a gateway that limits what the AI can reach, each with an audit trail.

Cost and abuse limits

Every user, retry and bad actor costs you money. I put the limits in before launch, with a budget per customer, caps on what each tool call can pull back, sign-up that's closed or verified, and the cheapest model that does each task well.

Safe retries

Agents retry, time out and get redeployed in the middle of a task. I build each step to be safe to run twice, with durable workflows and deduplication keys, so a retry never writes a record twice or charges a customer twice.

Dependable results

A model's output varies from run to run, so I treat it as a draft until something checks it. I check tool calls against a schema and answers against real data, send the important ones to a person, and only ship a new model that beats the current one.
Customer AWhat are myopen invoices?Customer BWhich of my productssell best?Customer CA script, on repeat:Summarize all myinvoices again.C’s spendbudget reachedLimits, per customer2AI assistantwrites a query for each questionPermissionsoutside the model1Company databaseCustomer A’s dataCustomer B’s dataCustomer C’s data1 · Security and permissions2 · Cost and abuse limits

From client work

All case studies →
A secure remote MCP server for database accessStaff point AI agents at production data through one hosted server behind company sign-in, with permissions enforced outside the model.
Access setupLocal → HostedIoT sensor platform · In production · internal
→
Infrastructure for a production vision modelLLM labeling, versioned training data and a promotion gate behind a custom vision model, now run by the client's own team.
Labels wrong: frontier models → this model15–20% → ~4%IoT platform · In production
→

Questions

We want to let an AI agent query our production data. Is that safe?

It can be, if the limits sit outside the model. The server authenticates the caller and checks authorization before a query reaches the database. Use read-only credentials with access limited to the data that caller may see, bound the query's runtime and output, and log who asked. People with different rights need separate authorization boundaries. A prompt is not an access control.

Can't we just use an off-the-shelf MCP server?

For one developer on their own machine, often yes. MCP, the Model Context Protocol, is how AI assistants reach tools and data. A ready-made server runs queries under the login it receives. Across a team, local installs can drift and each device needs its own setup. A hosted MCP server centralizes that setup and can use company sign-in, with authorization enforced by the server and database.

Should we run AI through Bedrock or our own hosting, instead of signing up with Anthropic or OpenAI directly?

It depends on what you need to control. Going through a cloud you already use, such as Claude on AWS Bedrock, keeps the model inside the access controls, region and billing your company already has, which often settles a security review. Going direct to a lab usually gets you its newest models and features first. Hosting open-weight models yourself, or on a host like Baseten, makes sense when the weights must stay with you, the volume is steady and high, or you want to tune a model for one task. I start from the data you’ll send and the review it has to pass, then pick the simplest setup that passes it.

Do we need the most expensive model for everything?

Rarely. The frontier models earn their price on hard reasoning and open-ended questions; sorting, extracting and summarizing usually run as well on a model that costs a fraction as much. I choose per task, by testing the candidates on your own examples, held out from anything a model learns from. Sometimes the best choice is a task-specific model: one IoT platform’s own model gets close to 96% of labels right on its task, where general-purpose frontier vision models average 80 to 85%.

Models keep getting cheaper. Why does our AI bill keep going up?

Because cheaper calls get made more. As each call costs less, features call the model more often, prompts carry more context, and agents loop and retry, so total spend climbs; economists call it the Jevons paradox. Set limits before launch, outside the model: a budget per customer, bounded tool runtime and output, controls on account creation, and a model chosen for each task. Query limits belong in the server and database rather than the prompt.

Tell me what you're building

A free 30-minute call with me, not a sales team. Bring the diagram or the prototype. My engagements run long and personal, so I'd rather know we fit first. Either way, you'll get a Slack Connect invite for later questions.

Book a call

How an engagement runs

First callFree, 30 minutes. You tell me what's in the way, and I'll tell you whether it's a problem I'd take on. It's a fit only if we both say so.
ShapeEngagements flex to the problem: an assessment, a build, a rescue, or working inside your team. We pick the shape after the call, not before.
DeliveryEverything ships through your repos and your review process, with the tests and docs your team needs to run it without me.