AI Dev Tool Radar

Inspiration

The AI tooling ecosystem moves faster than anyone can read — releases, breaking changes, repo drama, HN debates. Analysts are slow and rarely cite sources; LLM summaries hallucinate. I wanted the thing I'd actually use: an analyst that never sleeps and always shows receipts.

What it does

AI DevTool Radar is a fully autonomous analyst. My only input is what to watch: GitHub repos and blogs/news feeds (OpenAI, Anthropic, DeepMind, Hugging Face, Hacker News by default). Every source becomes a real Airbyte pipeline syncing hourly into ClickHouse. On a schedule, the agent writes its own SQL against the live data, fetches the full source articles and HN threads for what it selects, drafts a briefing where every claim carries an inline citation, grounds the evidence in Senso, and publishes to cited.md — with every run traced in Langfuse (~$0.0002 each).

How I built it

Each sponsor tool is a whole layer, not a checkbox: Airbyte is the ingestion layer and the control plane (the app creates/deletes pipelines via its API when sources change), ClickHouse is the memory and analysis engine, Senso/cited.md handles grounding and publishing, Langfuse observability, and Render hosts the full stack — dashboard, API, and scheduler in one service. The agent core is ~300 lines of Python: schema in, model-written SQL out (behind a SELECT-only guard), briefing published.

Challenges

• Senso's publish path wasn't documented — I reverse-engineered their CLI's npm package to find the content-engine endpoints, and unlocked publishing via an undocumented destination-activation call.
• ClickHouse Cloud quirks: no FINAL on SharedMergeTree (the agent dedupes with argMax), stateless HTTP sessions.
• OpenAI/Anthropic block bot RSS readers — swapped in Google News mirror feeds; my reconciler auto-created the replacements and deleted the dead pipelines.

What I learned

Give the model the schema and let it write the queries. SQL against ground-truth data beats free-form summarization — every claim traces to rows. And using one tool deeply beats ten shallowly.

What's next

Premium deep-dives behind an x402 paywall so other agents pay to read them, and a critique-and-revise pass to sharpen the briefings.

Built With

• airbyte
• clickhouse
• github
• hacker-news-(algolia)-api
• langfuse
• managed-via-its-rest-api)-database:-clickhouse-cloud-ai:-openai-api-(gpt-4o-mini)-grounding/publishing:-senso-api-?-cited.md-observability:-langfuse-hosting:-render-(web-service:-frontend-+-backend-+-scheduler)-other-apis:-github-rest-api
• radar
• senso