Information has never been more abundant. Intelligence has never been more rare. The gap between the two is the only thing I am interested in building.
The gap
Most of the knowledge that would actually help you make a decision is sitting in a form you cannot query. It is spread across PDFs, web pages, reports, inboxes, and exports. It is real signal, and it is effectively invisible, because nothing can search across it and give you an answer you can trust.
You can feel the cost of this every time you know the answer is "in there somewhere" and you still have to spend an afternoon finding it. Multiply that across an organization and it is no longer an inconvenience. It is a tax on every decision you make.
Not a chatbot
The reflex right now is to point a chatbot at the problem. Paste in some text, ask a question, accept whatever comes back. That is fine for drafting. It is not infrastructure, and it is not intelligence, because you cannot see where the answer came from.
An intelligence engine is different in one decisive way: every output carries its sources. It does not just generate a plausible paragraph. It retrieves the specific passages that support a claim, cites them, and lets you verify. If an answer cannot be traced back to a source, it does not count.
A demo impresses once. Infrastructure compounds. The difference is whether you can trust the output enough to act on it.
Why infrastructure comes first
It is tempting to start with the flashy vertical, the thing with an obvious buyer. I went the other way on purpose. The hard, reusable part is the engine underneath: ingestion, structuring, retrieval, and sourcing. Build that once, well, and every product you put on top inherits it.
That is the whole thesis in one line: turn fragmented information into structured, searchable, actionable intelligence, and make that a foundation rather than a feature.
The layers
I think about the system in five layers. Each one turns the layer below it into something more useful.
- Acquisition. Crawlers, ingestion pipelines, parsing, metadata extraction. Get the raw material in.
- Semantic structuring. Embeddings and a vector store, so content can be addressed by meaning, not just keywords.
- Context memory. Track sources and entities so provenance and context survive across queries.
- Reasoning. Synthesis across multiple sources, not a single lookup.
- Delivery. Search, reports, and sourced answers through an interface a stranger can navigate without a manual.
Foundation, then verticals
Concretely, that becomes a sequence. Manifest is the foundation: the document intelligence engine that handles acquisition, structuring, retrieval, and sourcing. Aletheia is the reasoning layer that sits on top and turns retrieval into analysis. Verticals like Playbook aim the whole stack at a specific market.
The verticals are where the value becomes obvious to a buyer, but they are not where the work starts. They start as direction. They ship as products only when they are ready, and not a day before.
What I am optimizing for
Three things, in order. Sourced, so every answer can be verified. Structured, so the corpus is addressable instead of a pile. Demoable, so someone who has never seen it can sit down and understand it in a minute. If a feature does not move the system toward gathering, structuring, reasoning over, or delivering intelligence, I question whether it belongs.
That is the bet. Not a smarter chatbot. A foundation that turns scattered information into an advantage, and a set of verticals that prove it, one market at a time.