The platform
Beacon is our modular, tool-based platform — orchestration, grounding, evaluation and delivery, already engineered. We compose the parts and fit them to a client's workflow, so a bespoke analyst stands up in weeks, not years. It's the scaffolding; the domain expertise is the only custom layer.
The brief · Basis Global
Basis Global runs continuous brand trackers for clients like Habitat and Sainsbury's. Turning that survey data into a decision a CMO will act on takes a scarce senior analyst — and a deck where every number survives scrutiny.
So we became the analyst
The data and the tools are commodity; the judgment is the moat. We embedded with Basis's experts and turned three things they carry in their heads into code the agent runs on every question.
Before it can answer · data discovery
There's no pre-built ETL and no warehouse modeling per client. The agent profiles every column in SQL, reconciles the spec's question codes to the real columns, derives the breakouts, and writes its own catalog — so research runs in natural language against whatever you uploaded.
The analyst, running
Orient, plan, investigate, assess — the agent routes across Claude models (Sonnet to plan, Haiku for fast calls, Opus to synthesize) and pulls survey data through MCP tools, exactly as the analyst would. A live trace of the real pipeline, not a mock.
The research loop
The loop doesn't stop at the first answer. It judges its own coverage twice — a deterministic gate, then a senior-analyst reflection that lists gaps by severity. Closeable gaps reopen the loop and target the unused metrics; uncloseable ones become honest caveats. All of it bounded by hard caps.
One question, an investigation
From one question the agent opens several investigation angles, expands each into hypotheses, and grounds every one in weighted SQL on DuckDB — significance-tested, then synthesized into a cited answer. Hover any node to see the exact query behind it.
True by construction
Every figure carries a quiet evidence marker — open it for the proof: the measured value, its base and honest significance, the exact survey question, and the SQL behind it. Nothing floats free of its claim, and an untied number never reaches the deck.
The story, not the dump
An analyst doesn't hand you charts — they build a case. The agent writes the thesis first, sequences the evidence to land it, and demotes what's tangential to the appendix — every section tied back to a research package that proves it.
Inside the storyboard
From your business question it derives the story spine, verdicts what the data already supports, then asks before it researches the gaps. The result is a living storyboard of flowing headlines — each carrying how well it's backed: done, a research gap to close, or a judgment call left for you.
Shipped, not demoed
The storyboard compiles to a client-branded PPTX with no LLM at render time, gated by the Basis Scorecard before it ships. This isn't a prototype: it's deployed, evaluated, and regression-tested.
The pattern
Basis is the worked example, not the limit. The engine is Beacon; only the domain layer — the expertise we embed to learn and encode — changes per client. Built to Anthropic's agent-design standards, on the most capable Claude models.