Agentic Discovery Platform

An AuroPro consulting accelerator. Point an LLM agent at an enterprise's messy business documents; it discovers the contradictions, undocumented processes, and control gaps — then produces a client-ready report suite. Runs on any domain with zero config.

_{Internal / proprietary · v1/ = the engine · research/ = market & UI analysis}

---

How it works

flowchart TB
    D["📂 Business documents<br/><small>SOPs · policies · RACIs · system exports · working notes</small>"]

    subgraph DISCOVER["① Discover — live agent loop"]
      direction TB
      A["Read &amp; classify each document"]
      T{{"Generic tools<br/><small>describe · group_by · join_diff<br/>filter_count · find_mentions</small>"}}
      X["Cross-reference<br/><small>documented process vs. actual data</small>"]
      A --> T
      T -- "code computes the numbers" --> X
      X -- "more evidence needed" --> T
      X --> F["Findings<br/><small>contradictions · gaps · unowned processes</small>"]
    end

    subgraph REVIEW["② Verify"]
      G{"Grounding gate<br/><small>every number traces to a tool result</small>"}
      H["👤 SME review (HITL)<br/><small>accept · override · route</small>"]
      G -- "ungrounded → rejected" --> A
      G -- "grounded" --> H
    end

    subgraph DELIVER["③ Deliver"]
      S["Synthesis<br/><small>pain points · opportunities · roadmap</small>"]
      C["Confidence + provenance tiering<br/><small>cite source docs, business language only</small>"]
      R["📑 6-report client suite<br/><small>HTML + Markdown</small>"]
      S --> C --> R
    end

    D --> A
    F --> G
    H --> S

    %% high-contrast: dark text on light fills (readable in light & dark themes)
    classDef src     fill:#dbe4f3,stroke:#33425b,stroke-width:1.5px,color:#11161f;
    classDef agent   fill:#cfe0ff,stroke:#1f4fa3,stroke-width:1.5px,color:#0a1f44;
    classDef gate    fill:#ffe6b3,stroke:#9a6a00,stroke-width:1.5px,color:#3d2a00;
    classDef human   fill:#e8d8ff,stroke:#6a3bb0,stroke-width:1.5px,color:#2a124d;
    classDef out     fill:#c9efd6,stroke:#1b7a45,stroke-width:1.5px,color:#0a2f1a;

    class D src;
    class A,T,X,F agent;
    class G gate;
    class H human;
    class S,C,R out;

Loading

The agent reasons; code does the math. It loops over deterministic tools to compute over the data, so every figure is exact and traceable — never an LLM guess. The grounding gate bounces any ungrounded number back to the agent; an SME reviews findings before they become a deliverable.

---

The output: a 6-report suite

#	Report	Role
01	Current State Assessment	Factual baseline + process-flow diagram (no judgements)
02	Pain Points & Opportunities	Issues ranked by impact, each → an opportunity
03	Transformation Recommendation	Value vs. feasibility 2×2 + sequencing
04	AI Opportunity Portfolio	Centrepiece — before/after per opportunity
05	Transformation Roadmap	Three horizons
06	Supporting Artefacts	System/data-flow map + source-document index (audit trail)

---

Design decisions (the why, in one line each)

• Code does the math, the agent reasons → numbers are exact & consistent across runs, never hallucinated.
• Grounding gate → a finding's every number must trace to a tool result, or it's rejected.
• No platform jargon in client output → a leak guard fails the build if tool names/filenames slip in.
• Confidence by provenance, not self-report → findings cite the documents they rest on.
• Live by default, any domain → report content is generated per run; an O2C fixture exists only as a demo-safe fallback.
• Client-agnostic → no org name unless detected from the docs; a run can suppress it.
• Human-in-the-loop SME review → the trust step and the basis of the "less stakeholder time" story.

Full rationale: v1/docs/ · competitive positioning: research/.

---

Quickstart

First time on this repo? Two commands from the root — no API key, no cost.

Prerequisites: uv (Python env manager) and Node.js 20+ (npm). The setup step tells you if either is missing.

macOS / Linux:

make setup      # Python env (v1/) + explorer deps
make run        # OFFLINE demo: golden replay → builds the explorer → opens the report suite

Windows (make isn't native — use the cross-platform runner; identical behaviour):

python tasks.py setup
python tasks.py run

run is a deterministic golden replay (the saved run, rendered offline), so it needs no credentials and can't hit the "no cached response" wall. Run make (or python tasks.py) with no target to list every task.

Want a live run (real agent, spends credits)? Add a key first:

# put ANTHROPIC_API_KEY=… (or the AZURE_OPENAI_* vars) in v1/.env — setup seeds it from the template
make doctor     # (or: python tasks.py doctor)  — verify the provider is reachable (one tiny call)
make live       # the real pipeline — minutes, costs credits

A live run preflights your credentials: if v1/.env has no real key (or still has the .env.example placeholder), it stops immediately with a one-line fix instead of a traceback or a misleading "use the golden run" message. With no key, use make run for the full offline demo.

Interactive Console — the 6-stage operator flow with both servers running and the explorer embedded (make console / python tasks.py console). It starts the backend (:8742) and the explorer dev server together, opens the app in your browser, and keeps both running until you Ctrl-C (which stops both). No key needed — if no report data exists yet it renders the offline golden suite first; a live run is then a choice inside the Console UI.

Run on any domain by dropping its documents in v1/inputs/<domain>/, then make run DOMAIN=<slug> (or python tasks.py run --domain <slug>).

Manual equivalent (no make / no tasks.py)

cd v1
uv sync                                   # env + deps
cp .env.example .env                      # add ANTHROPIC_API_KEY (or Azure vars) for live runs
uv run python run.py --domain o2c --golden --auto-resolve   # offline demo → opens out/o2c/index.html
uv run python scripts/doctor.py           # (live only) check connectivity
uv run python run.py --domain o2c --auto-resolve            # live run (needs a key)

Develop

uv run pytest                                   # tests
uv run pyrefly check discovery run.py scripts   # type-check product code
pre-commit install                              # enable hooks

main is protected — changes go through a PR with passing CI. See CONTRIBUTING.md.

Layout

v1/      discovery engine + 6-report suite, tests, scripts, design docs
research/        market/competitor dossiers + UI/UX research (+ screenshots)
shared_context/  received engagement material (local only, gitignored)