Skip to main content
All work
Status: Led At Oracom Intelligence

Oracom AI platform: technical direction, team leadership, hands-on engineering

From the first stack decisions to leading Oracom's IT team: what I led, what I implemented myself, what I designed for colleagues to build, and what is still in progress.

  • TypeScript
  • Nuxt
  • Tailwind CSS
  • Node.js
  • NestJS
  • PostgreSQL
  • AWS
  • Agentic AI
  • RAG
  1. Step 1: A situation comes in
  2. Step 2: The matching flow, defined in the builder
  3. Step 3: The AI acts as the flow says
How the agent flow builder shapes the assistant's behaviour. Conceptual, no internals.

Context

When I joined Oracom GmbH in 2024 it was a call center. Since then it has grown into a company that builds AI communication products: voice and chat bots, smart assistants, and mail automation. The platform behind them was built from the ground up, and engineering work on it is ongoing.

My role

I have set the platform's technical direction since the first stack decisions, and I write production code. Since July 2026 I lead Oracom Intelligence, the company's IT team of five team members and me: I set priorities with the CEO and other departments, guide architectural decisions, review implementation, and stay hands-on in reliability and integration work. The platform is team work, so the list below keeps apart what I led, what I implemented, what I designed for others to build, and what is still open.

What I built and decided

  • Led: the original tech-stack choice and platform architecture.
  • Led: priorities agreed with the CEO and other departments, with unclear or urgent requests clarified and ordered before the team commits to them.
  • Led: staged plans that make ownership and dependencies explicit, for example a knowledge-ingestion feature planned as service work first, then backend integration, then UI, reusing existing capabilities and with acceptance criteria for quality, latency, reliability, and cost. That one is a plan, not a shipped result.
  • Led: decision gates before bigger commitments, such as a reproducible end-to-end proof before a larger third-party integration, and a baseline plus a representative pilot before changing testing tools. Both are defined, neither has a result yet.
  • Built with colleagues, then refactored by me: an agent flow builder that defines how the AI handles specific situations. After the first version I did its complete refactor, fixed bugs, and added checks before a flow is published.
  • Implemented: lifecycle handling for real-time conversations, with generation tracking, cancellation, and cleanup in the backend and WebSocket connection handling in an avatar app.
  • Implemented: the refactor of the delivery pipeline for on-site assistant devices.
  • Implemented: a generated, versioned API client package from the backend, adopted by several services in place of their own duplicated client code.
  • Implemented: structured logging and then OpenTelemetry tracing across the core services.
  • Implemented: mail automation, from processing incoming mail to sending templated messages.
  • Improved or mentored to completion: internal operational tooling such as a support dashboard, with organisation and contact lookup, protection against stale results overwriting newer ones, server-side restrictions on privileged admin actions, and avatar management.
  • Improved or mentored to completion: integration work such as streaming speech-to-text.
  • Designed and reviewed: the document upload, embedding, and retrieval pipeline, built by a teammate I coached.
  • In progress: a staged infrastructure programme, starting with observability and small migrations that can be rolled back.

Outcome

The platform exists and is under active development. This page makes no claims about usage, customers, or revenue. What it shows are decisions and changes: the architecture the platform started from, the parts I built, the work colleagues built from my designs, and plans with explicit criteria that still have to prove themselves.

What I'd do next

Turn the planned gates into results: finish the first stages of the infrastructure programme, run the testing-tool pilot against its baseline, and add interaction tests for real-time conversations so reliability is measured rather than assumed.