Build log · Applied AI

Build log: small applied-AI systems

What happens to an AI system once somebody has to operate it — and what I am building to find out.

Portrait of Brenda Manrique
5 min read· Updated
Prototype and validation stage

This is a build log of the applied-AI projects Brenda Manrique has been building on her own since moving to Berlin. It is hands-on project work to learn what it takes to run small AI systems for real, not a business: she does not offer services through this site. Most of it is prototypes and infrastructure; one system is in production — the chat agent on this site.

The question#

The interesting part is not whether a model can call a tool. It is what happens next: deployment, permissions, failure recovery, updates, and the boundary between an agent and the systems it is allowed to touch.

Those are ordinary software questions. They are also the ones that separate a demo from something a team can run on a Tuesday. Answering them takes building, so that is what this is.

The kind of problem#

Repetitive workflows where people spend their time moving information between inboxes, WhatsApp, calendars, documents and internal tools. A useful system collects the right inputs, calls approved systems, keeps state, asks for approval when it matters, and leaves a trail.

Message / eventRouterTyped tools + dataApproval gateActionAudit / telemetry

The stack#

Python + FastAPI

Services she can read, test and deploy, rather than critical behaviour hidden inside a visual workflow tool.

Docker + VPS

Each deployment can be versioned and reproduced.

PostgreSQL / Supabase

Durable state where it is needed.

Retrieval + typed tools

MCP-style boundaries, so "know something" and "do something" stay separate.

WhatsApp interfaces

Most operational users are not going to live inside a new dashboard.

Monitoring and approvals

Including the rollback path, because without one there is no support model.

PythonFastAPIDockerPostgreSQLSupabaseRAGTool callingHITL

A shape that repeats#

The goal is to avoid a future where every deployment is a mysterious script on a server. Configuration, knowledge, tools and policies are per-deployment and versioned. The runtime and telemetry are shared.

config/          # approved systems, business rules
knowledge/       # indexed, versioned sources
tools/           # allowed actions, typed
policies/        # approval + escalation rules
runtime/
  agent service
  state
  telemetry
ops/
  deployment version
  alerts
  rollback path

Where it stands#

The portfolio chat agent is the proving ground: a public, adversarial surface that forces retrieval, evaluation, security and observability to actually work. It is in production. The rest is prototypes and infrastructure.

Frequently asked questions#

Is this a consulting business?

No. It is a build log of her own projects, a way to go deep on applied AI. She does not offer services or take on clients through this site. Most of it is prototypes and infrastructure; the portfolio chat agent is in production.

When did this start?

The experimenting started after she moved to Berlin in 2025. Around March 2026 it became more deliberate, with a focus on how these systems are deployed, monitored and rolled back.

Is she looking for a job?

Yes. She is looking for her next senior software-engineering or applied-AI role. These projects are part of how she keeps her engineering sharp.

Let us talk

Open to senior software-engineering and applied-AI roles. Happy to talk about any of this.

Portrait of Brenda Manrique

Brenda Manrique

Senior Software Engineer · Full-stack, financial systems, applied AI

Senior software engineer in Berlin. Previously Moody's Analytics, JPMorgan Asset Management and Money.Net. Now building applied-AI systems independently.

More about the author →
© 2026 Brenda Manrique. All rights reserved.|Privacy|