Factory of Innovations and Solutions

Description AI-powered LangGraph agent that turns messy procurement intake (files/text) into a structured, evidence-backed procurement event payload. Problem statement Procurement requests arrive in inconsistent formats and channels, so teams spend manual effort deciding if a request is processable, extracting fields, choosing templates, and mapping complex data (like line items and questionnaires) into an API-ready schema. Building naive AI-based Agent leads to trust issues from users side because Agents tend to hallucinate, especially on huge files. Solution This Agent automates that pipeline end-to-end: ingest and parse files, perform vector-based retrieval (RAG), extract structured event modules (Header, Worksheets, Suppliers, Questionnaire), validate against business rules, and output a structured JSON payload alongside an interactive chat interface for user refinements. Trust issues are handled via citations, so user can always easily track down the source of the data which is populated by the Agent. How it works A LangGraph orchestration pipeline runs nodes in sequence: Ingest data (LlamaParse + Qdrant index) -> deep agent extracts of the contents of file -> researches the file structure and outlines the data to be retrieved -> extracts data and adds citations -> maps extracted data into internal representation of Event -> iterates with human for any required adjustments of resulting Event. The deep agent uses RAG tools and a virtual filesystem to extract data, running self-correcting loops against structure and business validators. If human judgment is needed, it pauses via an interrupt and waits for user input. A secondary graph enables interactive, LLM-driven editing of the generated draft with real-time thinking preambles. Why it wins Deterministic validation loops + strict JSON contracts + retrieval-backed extraction reduces hallucinations, improves consistency, and converts unstructured intake (even on high volumes) into actionable payloads with minimal human triage, backed by a seamless human-in-the-loop chat experience along with citations for making user sure that extracted data is correct. GIT repo: https://gitlab.fmdev.io/maksim.lahunovich/hackaton-fy-27-factory-of-i-and-s/ Folder with documentation: https://gitlab.fmdev.io/maksim.lahunovich/hackaton-fy-27-factory-of-i-and-s/-/tree/master/docs Initial Kano session log: https://gitlab.fmdev.io/maksim.lahunovich/hackaton-fy-27-factory-of-i-and-s/-/blob/master/docs/Initial%20Kano%20session.md Implemented features with descriptions: https://gitlab.fmdev.io/maksim.lahunovich/hackaton-fy-27-factory-of-i-and-s/-/blob/master/docs/Phase%200%20features.md Loom video with cute cat (kit) running around: ... P.S. Attaching our team photo from day 1 Kano session and photo taken just before the submission. We had no clue about amount of work we'll need to do in those 6 days... P.P.S. Some evals screenshots attached as well :) P.P.P.S We got cut off by Loom at 5 min point, so video is cut, but supplier was created ;)