Consularis.ai

Process intelligence for small businesses. Map operations in plain language with an AI agent, get a report on what to optimise and automate, no 5k–15k consulting bill. Powered by Amazon Nova.

Comment

Amazon Nova Hackathon — Consularis Submission

Text description (for Devpost form)

A brief summary of your project, its purpose, and how it leverages Amazon Nova foundation models:

Consularis uses Amazon Nova on AWS Bedrock so mid-sized businesses, in particular physical ones (pharmacies, local chains, retail), can get process intelligence and automation guidance without the usual consulting bill. These businesses are often stuck in repetitive routines without knowing there is a way out, and they are the ones that need automation the most. Auditing and process mapping alone typically cost $5,000–$15,000 (or $100–$250/hour), which puts that intelligence out of reach. An agent (Aurelius) guides them through process mapping in plain language and generates a Company Process Intelligence Report: key metrics, process landscape, and AI-written recommendations on what to automate and which tools to use. Nova powers the conversational editor, the automation analyzer, and the report narratives, so they get audit-grade insight and a shareable PDF. Built for the Agentic AI category: one reasoning agent, one workspace, one report.

Inspiration

Before any RPA or workflow automation goes live, someone has to answer: Which processes? Which steps? What’s the ROI? That auditing and process-mapping phase is where consulting and automation agencies charge heavily, $5,000–$15,000 for a first engagement, $100–$250/hour for auditing, and mid-market projects into the $20,000–$100,000 range. For mid-sized companies, in particular physical businesses (pharmacies, local retail chains, growing SMBs), not software firms, those numbers make automation consulting a non-starter. They simply can’t justify the upfront cost to even explore what to automate.

Yet these are often the businesses most stuck in repetitive work, manual data entry, rekeying between systems, the same routines day in and day out, often without knowing there is a way out. That grind makes the day-to-day boring and demoralizing, and it makes growth hard, hiring more people just scales the same inefficiency. There’s a huge opportunity in serving this segment: they need process intelligence and automation roadmaps as much as enterprises do, but the traditional consulting model leaves them behind. We wanted a tool that could give them audit-grade intelligence anyway, map your operations in a structured way, get a report that says what to automate and why, powered by a single AI agent, without the price tag that keeps mid-market out.

What it does

Consularis is a process intelligence web app with three pillars:

  1. Conversational process mapping (Aurelius)
    We give mid-sized, physical businesses a way to map their operations and get an actionable report without paying for traditional auditing. Users can also add nodes, edges, and draw the process graph themselves in a visual, intuitive way, but that’s slow and tedious. So we offer a conversational path: users work in a workspace (company/session), chat or speak with the agent powered by Amazon Nova on AWS Bedrock, and describe their processes in plain language. The agent handles both low-level requests (“delete this edge”, “add a node between X and Y”) and higher-level ones (“please delete this flow”, “it should look more like …”, “merge these two steps”). It has tools to read and update a hierarchical BPMN 2.0 graph (nodes, edges, metadata such as duration, cost, error rate, automation potential). The diagram updates live as the user and agent refine steps, owners, and automation notes. No BPMN expertise required.
  2. Automation analysis Once the graph is populated, the user can run an Analyze flow. A separate Nova call (analyzer) receives the full graph summary and returns markdown: which steps are strong automation candidates, which tools (e.g. n8n, Zapier, Power Automate) fit, and a short CTA to book an appointment with Consularis for implementation. All driven by Nova’s reasoning over the structured process data.
  3. Company Process Intelligence Report
    The Report aggregates computed metrics (totals, per-process stats, costs, automation distribution, top issues) and sends them to Nova to generate two narrative sections: an Executive Summary (overview, key findings, top recommendations with concrete numbers and process/step names) and Automation opportunities (high-potential steps, a proposed workflow, and next steps). We do cost and general analysis and compute metrics that are useful for general understanding and financial planning, not just for automation. The frontend renders this with charts (bar charts for metrics, pie charts for automation potential and current state) and supports PDF export for sharing with stakeholders.

How we built it

  • Backend: FastAPI (Python), Uvicorn. Frontend: React, Vite.
  • Data: Process data in an in-memory SQLite-backed store; BPMN 2.0 is the source of truth.
  • Nova in three places: the main chat agent with tool use for graph edits; the analyzer (single-turn, read-only) for automation recommendations; the report generator (two narrative sections, each with a dedicated prompt and metrics context). Bedrock Converse API with Nova foundation models (e.g. Nova Pro, Nova Lite).
  • Agent tools: constrained operations on the graph (get_node, update_node, add_edge, etc.) so Nova produces incremental, validated updates instead of raw XML.
  • Separation: the analyzer and report writer only consume metrics and produce text, which kept prompts and behavior clean.

Challenges we ran into

  • BPMN vs. AI-friendly process model. We first wanted to use standard BPMN graphs, but their file structure was not AI-friendly and the libraries were bad. So we came up from scratch with our own augmented process files that are easy for the AI to understand and data-augmented so we can compute the necessary analysis.
  • Keeping the report and UI in sync. The report depends on the current session graph and computed metrics. We had to ensure the frontend requested the report only when data was ready and that PDF layout (charts, sections, headers) behaved well across screen and print. We iterated on CSS (print margins, chart scaling, section breaks) so the exported PDF looked professional.
  • Tool-calling reliability. Nova’s tool use is strong, but we had to design tool schemas and error messages so that invalid or ambiguous user requests (e.g. “change step X” when X doesn’t exist) led to clear feedback and retries rather than broken graph state. Validation in the backend after each tool call was essential.
  • Balancing generality vs. domain. Our baseline BPMN and registry are tuned to a specific domain (e.g. library/consular-style processes), but the agent and prompts are written to be adaptable. We had to avoid over-fitting prompts to one vertical while still giving Nova enough structure (e.g. automation_potential, current_state) to produce useful recommendations.

Accomplishments that we're proud of

  • One agent, one workspace, one report. A single Nova-powered flow from “describe your processes” to a shareable Company Process Intelligence Report and PDF, exactly what mid-sized businesses need without the consulting price tag.
  • You can also speak to the agent. Users can chat or use voice to describe and refine their processes, not only type.
  • Structured graph editing that works. Giving Nova discrete tools over a BPMN-oriented store gave us reliable, incremental updates and validation instead of fragile one-shot generation; the diagram and metrics stay in sync as the user and agent refine the model.
  • Report narratives that feel tailored. The Executive Summary and Automation opportunities sections use real process and step names and concrete numbers from the metrics, so the output reads like a real audit deliverable, not generic advice.

What we learned

  • Structured tools beat free-form generation for graphs. Letting Nova call discrete tools on a BPMN-oriented store gave us reliable, incremental updates and validation (IDs, schema) instead of fragile one-shot XML generation.
  • Separating “analysis” from “editing” simplified prompts and behavior. The main agent edits the graph; the analyzer and report writer only read metrics/summaries and produce text. That separation made it easier to tune each Nova use case and avoid mixed tool-calling and long-form writing in one flow.
  • Report quality depends on context shape. Passing compact, consistent metrics (totals, per-process, distributions, top issues) as text to Nova produced more consistent executive summaries and automation sections than dumping raw JSON. We learned to design a small “metrics context” format and to ask explicitly for concrete names and numbers in the prompts.

What's next for Consularis

  • RAG pipeline to enhance the agent. Use a RAG pipeline with clean, selected BPMN models so the agent can leverage best-practice process patterns and give even better recommendations.
  • Richer graphs from conversation. We’d like the model to build full, more complex graphs with more complicated steps (branching, subprocesses, richer metadata) just through discussion with the user, so mid-sized teams can go from “here’s how we work” to a complete, detailed process model without touching the diagram by hand.
  • Persistent workspaces and collaboration. Move from in-memory sessions to persisted storage so teams can save, share, and iterate on process models and reports over time.
  • Tighter path to implementation. Connect the report’s automation recommendations to concrete next steps, e.g. one-click “request implementation” with Consularis or links to no-code/low-code tools (n8n, Zapier) so the report doesn’t just inform but accelerates action.

Built With

Share this project:

Updates