Claude · Gemini · Kiro · Langfuse · Qdrant · Ollama

Building with AI, in the open.

An automated AI news portal and podcast, an AI radar, tooling to build software faster with Claude and Kiro, and IoT experiments. This is where I put what I build and what I learn — everything here is running or open source.

News & podcastAn AI-news pipeline that publishes on its own.
AI radarA daily scan of the AI ecosystem, with opinions and verdicts.
Dev toolingRules, skills and hooks for Claude Code and Kiro.
IoTSmart-plug energy monitoring, in progress.

Tools I work with

Running on my own server, every day.

AI models & agents

ClaudeAnthropicGeminiOpenAIMistralOllamaHugging FaceNVIDIAElevenLabsWhisperLangfuseLiteLLMQdrantKiroGemmaLlamaQwen

Platform & data

JavaSpring BootPythonFastAPIVueTypeScriptPostgreSQLMongoDBRedisRabbitMQClickHouseMinIODockerCloudflareKeycloakKongNginxPrometheusGrafanaLokiSonarQubeWoodpecker CITerraformAnsibleGitHubTelegramSearXNG

Trademarks belong to their owners; shown only to indicate what I use.

01 — running

What runs every day

The news portal, the podcast and the radar share one pipeline. It is the project I learn the most from.

AI news portal

A pipeline collects AI news, summarises and translates it with LLMs, runs every summary through a critic-and-reviser pass, and publishes the portal and a newsletter — without a human in the loop.

Java 21 · Spring Boot · MongoDB · Qdrant · LiteLLM/Ollama on my own GPU · Langfuse

The podcast

The same stories turned into a daily show: script, text-to-speech, then a speech-to-text check of the audio to catch bad blocks, published as an RSS feed.

Text-to-speech · Whisper quality check · RSS

The AI radar

A daily job that follows the AI ecosystem: it extracts the entities mentioned in the news, snapshots them, gathers outside signals (Hacker News, GitHub), and has the LLMs write opinions and verdicts. Rules decide what may be published, and nothing goes out without a blind test.

Eight-phase daily job · LLM extraction, opinion and verdicts

02 — method

How I work with coding agents

AI agents write most of the code here. A written method decides what they may write, and hooks make that method mechanical instead of optional.

  1. 01 SPECEvery domain has a living spec: ALWAYS / NEVER invariants and acceptance criteria. Agents read it before touching code.
  2. 02 REDA failing acceptance test comes first, with at most three criteria per cycle. If it does not fail, it is not a test.
  3. 03 GREENAgents implement in clean contexts, from the written spec only, in parallel where the work is independent.
  4. 04 GATEFull test suite, quality gate, one build promoted by SHA, automatic rollback. Nothing is done until I can open it in production.
  5. 05 TRACEEvery LLM call is traced, every service measured, and operations go through read-only tools.
  6. 06 PANELA live panel shows what is running, what is done (with the commit) and what waits on me.

Every rule in the method records the dated incident that created it.

in use

Workflows

Small acceptance-driven cycles: at most three criteria, a plan written by one sub-agent from written material only, and tests and code written by other sub-agents that start with an empty context and see only the cycle file. The right model is chosen per task and switched by the orchestrator, never by me. Every session ends with a handoff file that the next one validates against git, and a live panel shows what is running, what is done (with its commit) and what waits on me.

Rules · skills · hooks · 309 archived cycles, as of Nov 2026

in use

Making Claude reliable

One standards repository, imported by every project through a fixed four-rule core and an index that loads the rest on demand. Living specs are injected before an edit, and a pre-commit gate refuses a change that leaves its spec behind. Commits and pushes go through one skill, destructive git is denied, and a daily integrity check finds broken pointers. Operations go through read-only MCP tools instead of SSH.

36 process rules · 18 stack rules · 72 skills · 6 hooks, as of Nov 2026

in use

Kiro harness

The same method ported to Kiro CLI, which bills per request rather than per token. Four agents with hard budgets (an executor on the cheapest model, a judge, a read-only reader and a one-call escalation), hooks that refuse unbounded file reads, block edits to the build harness and warn as the context fills, a file-based workflow engine that tells the agent the single next command, and a local vector index with a two-slot resume pointer.

4 agents · 15 hooks · 61 scripts, as of Nov 2026

03 — tools

Tools and the launch pipeline

What I made along the way, and how close each piece is to leaving the lab.

MCP tools

Read-only MCP servers that replaced shell sessions for looking at my own platform: 66 tools over containers, logs, metrics, pipelines and quality gates, with a tested redaction layer. Agents get typed tools instead of a terminal.

MCP · Prometheus · Loki · CI

Launch pipeline

Claude InsightsReal-time dashboard for Claude Code sessions.in progress · not released yet
mcp-ops66 read-only MCP tools over logs, metrics, containers, pipelines and quality gates.in progress · license and decoupling first
standarts-llmRAG and MCP server over my own standards, with context memory and model routing.in progress
doc-converterStateless document converter (PDF, Office, HTML, images to Markdown and back) with an MCP server.planned · the smallest to open
copilot-meeting100% local meeting transcription and summaries: a Mac app plus a GPU server.in progress · personal tool
infowhere-soundAudio generation with an automatic report of where the voice stumbled.planned

“Open source” means a public repository, a license, a history scan and an install that works on a clean machine. Nothing is announced before that.

04 — iot

IoT, from the wall socket up

The newest front, and still an experiment.

in progress

Energy monitoring

Smart plugs report hourly energy and power every five minutes. A collector stores the history in PostgreSQL and a small dashboard shows it. The next step is reading the energy bills from PDF and simulating tariffs, so the numbers can say whether a contract is worth keeping.

Python · FastAPI · PostgreSQL · Vue 3 · TP-Link Tapo P110

05 — book

The book

in production · Casa do Código

Desvendando a Mente Aumentada

A book for people who work with AI rather than build it: how the models understand language, what they do not know about themselves, and the language between the two — orchestration, context, prompts and the memory you build. It ends in practice, with specifications as a working language and your own programming agent. Fourteen chapters and two hands-on appendices.

In Portuguese · Casa do Código

06 — hello

Who is behind this

I’m Leandro Siciliano, a software architect in Brussels. InfoWhere.ai is my lab: the place where I test ideas on real systems, keep what works and write down why. Nothing here is a product for sale.

If something here is useful, or you spot something wrong, write to me. Tell me what you are building, or what you would like to see here.

Prefer email? show my email. I keep your message only to reply to you. No cookies, no trackers. Privacy →