PresentationExperiencesProjectsOpen SourceEducationPostsMusic Feed
GitHubLinkedInXSoundCloud

Getraline

Visit site

AI Engineer (apprenticeship) · 2024 to 2026

Stack
GoNext.jsReactPythontrpc-agent-goClaudePostgresS3DockerOpenTelemetry

The company

Getraline (1993) supervises collective equipment, primarily elevators, through connected boxes and the GetraNet SaaS platform. I worked in the R&D software team, owning the company's AI projects end to end.

Sintesia

The company's agentic platform, integrated into GetraNet: ask in plain language and the agent fetches company data, analyzes it like a data scientist, and delivers a PDF, DOCX or Excel report. The first version ran on a legacy graph-agent framework delivered by Coddity, hard to evolve and limited to answering. I made the call to rewrite it from scratch and designed the whole new architecture: a Claude Code-style agent in the cloud, where each session gets its own containerized Docker workspace (shell, filesystem, code execution) so the agent writes and runs code instead of computing in its head. Built on trpc-agent-go with a minimal stateless Go API (chi, AG-UI), multi-provider LLMs, large tool outputs offloaded as S3-backed artifacts, two-level memory, session summarization, sub-agent forks, reusable skills, a managed container pool, prompt caching to cut inference cost, and OpenTelemetry tracing to Phoenix.
Sintesia architecture overviewGetraNet 4 · TUIchat UIAG-UI · streamingGo API · chistateless · ×N replicasagent harness · trpc-agent-goAgent loopClaude · multi-providerContextmemory · summarizersub-agent forksWorkspaceDocker · shell · FSruns PythonTools & skillsGetraHub dataPDF · DOCX · XLSXContainer pool + registrycreate · reattach · destroyPostgreSQLsessions · memory · metadataS3artifacts · files · volumesOpenTelemetry traces → Phoenix · cost & perf on every run
The agentic harness, from the interfaces down to storage.

Getranalyse

Rewrote a costly, unreliable Python tool into a modular Go platform for the operations team, where each tool is an independent module on a shared core with a unified Next.js frontend. Alarm Analyzer checks elevator emergency-call tests cycle by cycle from box logs and call audio (DNSMOS speech quality on ONNX runtime, transcription), with AI deliberately kept to extraction and summary. Log Analyzer queries many servers in parallel and rebuilds a single colored timeline.

Open source

Contributor to trpc-agent-go, Tencent's Go agent framework that Sintesia runs on: artifact storage, persistence components, and per-session containerized workspaces. I upstream the pieces I then use in production.

Explorations

Proofs of concept that shaped the year's architecture choices: AI extraction of technical specs from elevator audit reports, an autonomous Bash/DevOps agent operating on the server infrastructure, and audio analysis of hotline calls.

Internal enablement

Helped roll out the Claude suite across the company: training teams on AI-native development tools, running workshops, sharing best practices, and tech watch on the agentic ecosystem.