A pipeline, not a chatbot.
DeepPolicy is a ten-stage multi-agent workflow purpose-built for legislative and fiscal documents. Each stage has a defined role, a defined output, and an audit trail that traces every figure back to its source.
Specialised roles, not one big prompt.
Each agent has a single job, a single role brief, and a single output schema. Analysts analyse; editors compile; sociologists segment. This is how a real research department works, and it’s how the pipeline mirrors one.
Numbers are not generated. They are computed.
Every aggregate figure in a DeepPolicy whitepaper is computed by a deterministic Python layer over structured JSON, not by a language model. LLMs do the reasoning; arithmetic does the arithmetic.
Every claim is auditable, or it doesn’t ship.
The deliverable is not just the whitepaper — it is the whitepaper plus the chain of citations connecting each claim to its source line. If you cannot defend a number in a board meeting, you should not be publishing it.
Inside the pipeline.
Each stage runs against the output of the previous one. Stages 02 and 06 fan out across multiple parallel agents (one per chapter, one per demographic) with bounded concurrency. The full pipeline is resumable — an interrupted run picks up at the last completed stage rather than restarting.
- Stage 01Document ingestionParser
The pipeline accepts national budgets, draft legislation, regulatory consultations, and statistical annexes in their native formats — PDF, DOCX, XLSX, HTML. Documents are segmented into structurally coherent units: chapters, schedules, articles, line items.
- Stage 02Per-chapter analysisAnalyst agents (parallel)
A specialised analyst agent runs against each chapter. Each agent is briefed with your selected analytical framework and produces two outputs per chapter: a Markdown analysis with reasoning, and a structured JSON record of every classified line item.
- Stage 03Deterministic aggregationNumerical layer (no AI)
All numeric fields from the per-chapter JSON are summed, cross-checked against the source totals, and reconciled. This step is deliberately not AI — financial figures should not pass through a generative model unverified.
- Stage 04Whitepaper synthesisEditor agent
A senior editor agent compiles the per-chapter analyses into a unified whitepaper, preserving every figure, every cited line, and every chapter-level conclusion. The output reads like a single human-authored document, not a stitched summary.
- Stage 05Demographic identificationSociologist agent
A sociologist agent reads the whitepaper and identifies every affected demographic and occupational segment — by income band, region, sector, life stage, employment status. Output is a structured demographic map.
- Stage 06Targeted communication briefsCopywriter agents (parallel)
A copywriter agent generates a tailored brief for each identified demographic — what changes for them specifically, in their language, framed for their concerns. Useful for campaign comms, member newsletters, GR briefings.
- Stage 07VisualisationPlotting layer (no AI)
Plotly-backed charts are generated deterministically from the structured data. No hallucinated figures, no inventive labelling — every chart is reproducible from the JSON.
- Stage 08Translation & localisationTranslator + proofreader agents (parallel)
Optional. Every output document can be translated into the working language of the publishing institution, then proofread by a second model with native-speaker prompting before delivery.
- Stage 09Cross-document synthesisSynthesiser agent
When you supply a second document — a party manifesto, a coalition agreement, a regulatory commitment — the synthesiser agent produces a side-by-side reconciliation: where the budget honours the commitment, where it contradicts it, where it is silent.
- Stage 10Export & audit trailCompilation layer
Final outputs are compiled to PDF, Markdown, and structured JSON. Every claim in the whitepaper carries a citation back to its originating line in the source document. The audit trail is the deliverable, not an afterthought.
Five deliverables. One audit trail.
A single, unified analytical document covering every chapter of the source legislation, formatted for institutional publication.
Optional. A line-by-line comparison between the source legislation and a second reference document — manifesto, coalition agreement, regulatory pledge.
Every classified line item, every aggregate, every chapter summary, in machine-readable form. Plug into your own dashboards.
Reproducible, deterministic visualisations of the fiscal and structural findings, ready for whitepaper inclusion or press use.
One targeted brief per identified affected demographic, ready for member comms, campaign use, or GR distribution.
Every figure and claim in every output traces back to a specific line in the source document. The audit trail is the moat.
Run it against your document.
Discovery calls walk through how the pipeline would handle a specific document you care about, what the output looks like, and which tier fits.