Gumdrop

Work · the deep dive

One platform, proving the whole model.

We built and run robotics.press end to end — autonomous research, a sourced and cross-linked knowledge base, a metered API. It's our prime demonstration: proof the model produces real depth and quality, autonomously, at scale. Here's how it actually works.

From raw sources to intelligence Messy raw sources go into the machine and come out as structured, sourced, confidence-scored intelligence. raw sources messy the machine structured · sourced · scored intelligence
Flagship

robotics.press

An automated intelligence platform we built and run — covering the drone-warfare domain from 380+ sources. It researches autonomously, structures everything into a sourced, cross-linked knowledge base, and serves it to paying customers through a metered API. Proof the model produces real depth and quality, autonomously, at scale, for cents.

AutonomousProvenanceMetered API
View live →

How the data works

It's a graph, not a pile of articles.

robotics.press models its domain as connected entities — every fact filed against an ontology, a structured map of what matters in the field: its classes and a controlled vocabulary. The ontology is the schema of meaning, and it's where the client's domain expertise lives.

Incident
A real-world event — when, where, what happened, the outcome.
System
The system involved, and its class.
Actor
Who acted, on each side — states, units, groups.
Manufacturer
Who makes the system.
Component & Supplier
What the system is built from, and who supplies the parts — the supply chain.
Conflict
The wider theater an incident belongs to.
Impact
Casualties, damage, economic loss.

The value is the cross-linking. An incident isn't a standalone fact — it's connected to the system used → its manufacturer → its components → their suppliers; to the actors; to the conflict. That web lets you ask questions no single article could answer.

The knowledge as a queryable network Two incidents share a system and a conflict; the system links out to its manufacturer and supplier, and each incident links to its actors and impact. The shared structure is what makes the knowledge a network you can question. Actor Impact System Manufacturer Supplier Conflict Incident Incident
Two incidents share a system and a conflict. That shared structure is what makes the knowledge a network you can question — which makers, which suppliers, which actors recur — not a pile of separate facts.

How it's enriched

The craft.

Raw reporting is messy, contradictory, unstructured. Enrichment is what turns it into the graph — four steps, each one auditable.

  1. 1

    Extract

    Pull out the structured facts — but only what a source explicitly states. Every fact points back to the exact words it came from. Nothing invented.

  2. 2

    Resolve

    Match messy mentions to canonical entities. Obvious matches are made deterministically; a model handles only the ambiguous residue, can link only to entities that already exist, and must clear a confidence bar.

  3. 3

    Stamp

    Every link records how it was made, the raw text behind it, and a confidence score. Auditable end to end.

  4. 4

    Cross-link

    Connect the resolved entities into the graph — where the value is.

A wrong link is worse than an honest gap.

Better to leave something unconnected than assert a connection that isn't real — because the whole value is trust.

The result is a record that's sourced, linked, and confidence-scored — far beyond the raw text it came from.

Want something like this for your domain?

You bring the domain expertise. We bring the machine that turns it into an owned, queryable, sellable intelligence platform.

Book a call