How We Built Open Signal: An AI-Powered Publication That Runs Itself

Why This Exists

The intelligence industry has a structural problem. High-quality analysis — the kind that helps people understand what is happening in technology, geopolitics, and markets — is expensive to produce and almost always locked behind paywalls. Bloomberg Terminal costs over $20,000 per year. Stratfor, Gartner, and similar services charge thousands for subscriptions. Even quality technology newsletters have moved to paid models.

The result is an information asymmetry where well-funded organizations have access to clear, structured analysis while everyone else relies on a combination of free media (optimized for engagement, not understanding), social media (optimized for virality, not accuracy), and word of mouth. This gap matters. In a world increasingly shaped by AI, semiconductors, and geopolitical competition, understanding these forces should not be a luxury good.

Open Signal is an experiment in fixing this. The premise is simple: use AI to dramatically reduce the cost of producing structured intelligence, then give it away for free. No paywall, no premium tier, no “subscribe for the real analysis.” Every piece of content we publish is available to everyone.

This is a deliberately transparent account of how we built it, what works, what does not, and what we are still figuring out.

The Technical Stack

Open Signal is built on Astro, a modern web framework designed for content-heavy sites. We chose Astro for several reasons:

Content-first architecture. Astro’s content collections system provides structured schemas for different content types, type-safe frontmatter validation, and a clean authoring experience in MDX (Markdown with JSX components). This matters because our content pipeline needs to produce well-structured output that passes validation automatically.
Performance. Astro generates static HTML by default, with JavaScript loaded only where interactive components require it. Intelligence content does not need client-side interactivity — it needs to load fast, render cleanly, and be readable. Astro delivers this without configuration overhead.
Flexibility. We use Tailwind CSS for styling, MDX for content authoring, and standard web technologies throughout. There is no proprietary lock-in. Every piece of the stack could be replaced independently.

The site is deployed on Vercel, which provides automatic deployments from Git, edge caching, and zero-configuration HTTPS. The deployment pipeline is straightforward: push content to the repository, and the site rebuilds and deploys automatically.

Content Architecture

We publish five content types, each with a distinct structure and purpose:

Deep Signals: Long-form analysis pieces (1,500-2,000 words) that examine a single topic in depth. These are the closest analog to traditional intelligence reports.
Signal Briefings: Daily structured updates with 5 items, each containing a summary and analytical commentary. Designed for readers who want to stay informed in five minutes.
Signal Maps: Visual and structured explainers that map competitive landscapes, technology stacks, or market dynamics. Heavy use of tables and structured data.
The Long View: Weekend deep dives (3,000-4,000 words) on structural questions that unfold over years, not days.
Open Source: Meta-content about the project itself — how it works, what we are learning, and what we are building next.

Each content type has a defined schema in Astro’s content configuration. This means every article must include specific metadata — title, description, publish date, tags, topics, and sources — and the build process validates this automatically. Malformed content fails at build time, not in production.

The AI Content Pipeline

This is the part that makes Open Signal economically possible — and the part that requires the most intellectual honesty.

AI language models can produce fluent, well-structured text at a marginal cost approaching zero. This is the enabling technology behind the entire project. Without it, producing daily briefings, weekly analysis, and monthly deep dives would require a team of analysts that we cannot afford to pay — certainly not for a free publication.

But fluency is not accuracy, and structure is not insight. The AI pipeline is a tool, not an autonomous author. Here is how we handle the tension between automation and quality:

Factual grounding. Every factual claim in every article must trace to a verifiable public source. The content pipeline uses publicly available information — earnings reports, product announcements, regulatory filings, published research — as source material. We do not fabricate quotes, invent statistics, or present speculation as fact.

Analysis labeling. We explicitly distinguish between reporting (what happened) and analysis (what it means). When we offer an interpretation, it is framed as an interpretation. When we state a fact, it is verifiable. Readers should always be able to tell which is which.

Uncertainty acknowledgment. When we do not know something, we say so. When predictions are uncertain, we state the uncertainty. The intelligence industry’s worst habit is false confidence — presenting probabilistic assessments as certainties. We try to avoid this.

Source attribution. Every article includes a sources field in its metadata, identifying the categories of sources used. This is not the same as inline citations in an academic paper, and we do not pretend it is. But it provides transparency about where the information comes from.

What AI Does Well Here

Structured synthesis is the AI pipeline’s strength. Given a set of facts about a topic — company earnings, product announcements, market data — language models are remarkably good at organizing that information into a coherent narrative, identifying the key dynamics, and explaining why they matter. This is the core of intelligence work: not gathering the information (which is increasingly public), but structuring it so that a reader can understand it quickly.

The pipeline also excels at consistency. Every Signal Briefing follows the same structure. Every Deep Signal has an identifiable analytical framework. This consistency is a feature, not a bug — readers develop expectations about format that reduce cognitive overhead and help them extract information efficiently.

What AI Does Poorly Here

Original reporting. The pipeline cannot call sources, attend conferences, or observe events firsthand. It works with publicly available information, which means it is always at least one step removed from primary sources.

Contrarian insight. Language models are trained on existing text, which gives them a tendency toward consensus views. The most valuable intelligence is often the analysis that contradicts conventional wisdom — and AI systems are structurally inclined toward conventional wisdom. We compensate by deliberately prompting for alternative interpretations and challenging assumptions, but this is an ongoing limitation.

Judgment under genuine uncertainty. When the facts are ambiguous and reasonable people disagree, language models tend to hedge rather than commit to a position. This is sometimes appropriate (genuine uncertainty should be stated as such) and sometimes a weakness (readers value clear assessments, even when qualified).

The Philosophy

Open Signal is built on three principles that guide every decision:

Radical Accessibility

Intelligence analysis should be free. Not “free with ads that compromise your attention.” Not “free tier with the good stuff behind a paywall.” Free. The entire publication, every article, every analysis, every data visualization — accessible to anyone with an internet connection.

This is economically possible because AI reduces production costs to near zero. It is philosophically motivated by the belief that understanding the world should not be rationed by income.

Transparency

We tell you how the sausage is made. This article is the most literal expression of that commitment, but the principle extends to everything we do. Our content schemas are public. Our source categories are listed on every article. Our limitations are stated openly.

Most publications ask you to trust their brand. We ask you to evaluate our methods.

Intellectual Honesty

We would rather be clearly wrong than vaguely right. Specific, falsifiable claims are more valuable than hedged platitudes, even when the specific claims sometimes turn out to be incorrect. When we make a prediction and it fails, we will say so — and explain what we got wrong.

What We Are Still Figuring Out

This is an experiment, and experiments have open questions.

Quality control at scale. As we increase publishing frequency, maintaining factual accuracy becomes harder. We are developing validation workflows that check claims against source material, but this is an area where the process is still evolving.

Audience feedback loops. A publication without readers is a journal. We need mechanisms to understand what our audience finds valuable, what they find lacking, and what topics we are missing. Building this feedback infrastructure is a priority.

Sustainability. “Free” is not a business model. Open Signal runs on minimal infrastructure costs today, but if the audience grows, so do hosting costs, content production demands, and the need for more sophisticated tooling. We are exploring options — sponsorships, grants, community funding — but have not settled on a model.

The evolving AI pipeline. Language models are improving rapidly. Techniques that work today may be superseded by better approaches in months. The content pipeline is designed to be modular — we can swap components without rebuilding the whole system — but keeping up with the pace of improvement is a perpetual challenge.

An Invitation

Open Signal is, at its core, a bet: that the combination of AI-powered production, open access, and intellectual honesty can produce something genuinely valuable. Not a replacement for deep human expertise, but a complement to it — a free, always-available layer of structured intelligence that helps people understand the forces shaping the world.

We do not know if this bet will pay off. But we think it is worth making.

The code is open. The content is free. The methodology is transparent. If that interests you, read the analysis. If it does not, that is fine too. The whole point is that it is here for anyone who wants it.