Sybil — AI Business Requirements Document Generator

Overview

Sybil is an AI-powered platform that ingests fragmented project communication (emails, meeting transcripts, slide decks, spec docs, spreadsheets — 36+ formats) and autonomously generates a structured 13-section Business Requirements Document with inline citations, conflict detection, and stakeholder sentiment analysis.

Built as a GDG Hackathon 2026 submission (Feb 21–22, 2026) in a 2-day sprint.

• Live demo: gdg-hackathon-brd.vercel.app/dashboard

What it does

• Multi-format ingest — drag-and-drop upload of PDF, DOCX, PPTX, XLSX, CSV, HTML, TXT. Parsing backed by my open-source Chomper library (forked and extended for this project).
• Agentic BRD generation — a Gemini 2.5 Pro agent autonomously decides its own workflow: reads document summaries, retrieves full text, runs cross-document searches, then emits 13 BRD sections (executive summary, objectives, stakeholders, functional + non-functional requirements, scope, risks, dependencies, success metrics, timeline, cost-benefit) with inline citations.
• Natural-language refinement — select any text in the BRD viewer, describe a change (“make this more concise”, “add acceptance criteria”), and the AI generates an edit. One-click accept.
• Conflict detection — the agent flags contradictions across source documents and offers a “Resolve with AI” action with cited fixes.
• Unified AI chat — single panel for Q&A, refinement, generation, and triggering full BRD runs. The AI self-classifies its response type so the UI adapts (edit bar only appears for editable output).

Architecture highlights

flowchart LR
    A["Project artefacts<br/>emails · transcripts · specs"] --> ING["Ingest + Parse<br/>(Chomper)"]
    ING --> EMB["Embeddings<br/>semantic filtering"]
    EMB --> AGENT["Gemini 2.5 Pro Agent<br/>plans its own workflow"]
    AGENT --> SEARCH["Cross-doc search<br/>+ conflict detection"]
    SEARCH --> AGENT
    AGENT --> BRD["13-Section BRD<br/>inline citations"]
    BRD --> UI["Next.js viewer<br/>natural-language refinement"]

• Frontend: Next.js 14 (App Router) + TypeScript, shadcn/ui, Tailwind, Zustand for state, Framer Motion. Deployed to Vercel.
• Backend: FastAPI (Python 3.11), fully async. Deployed to Google Cloud Run, infrastructure managed via Terraform (Artifact Registry, Cloud Build).
• Data: Firestore (NoSQL) for project/BRD metadata, Google Cloud Storage for raw documents.
• AI: Gemini 2.5 Pro for agentic reasoning, Gemini text-embedding-004 (768-dim) for semantic document filtering.
• Auth: JWT (HS256) with bcrypt password hashing.

Technical depth — the “virtual tool” pattern

The Gemini API doesn’t allow combining function-calling (tools) with structured JSON output (response_mime_type) in a single request — but this project needed both agentic planning and deterministic structured sections.

Solution: define virtual tools (submit_brd_section, submit_analysis, submit_response) as function declarations. The AI “calls” them inside its agentic loop — the backend intercepts each call and extracts the arguments as structured data instead of executing a real function. One request, agentic planning, typed output.

Other production-grade touches: fire-and-forget background generation (POST /brds/generate returns 202 while the agent runs up to 30 iterations / 65K-token turns); 3-layer prompt-injection defense (input validation + regex filters + defensive system prompts); full-text context instead of chunked RAG to preserve cross-document requirements.

Impact

Complete, deployed hackathon submission with 33 bundled sample documents across 3 realistic projects (e-commerce checkout, mobile auth, internal dashboard) showing the system correctly surfacing contradictions between stakeholder communications. Demonstrates end-to-end agentic AI in production: real cloud deployment, async backend, IaC, security hardening, and a polished UI — not a notebook demo.

Built with Vanshika as a 2-person team.