Brief AI — Agentic Packaging Design Brief Generator
Turns messy packaging inputs into a cited, reviewable design brief. Built for the Esko AIForge Hackathon 2026 (problem: From Product Idea to Design Brief).
Demo video: https://www.youtube.com/watch?v=y4wWjKvDw0U
Live demo: https://esko-brief-ai-m9fc.vercel.app (free tiers; if it hangs, please use the video)
The problem
A brand manager has an Excel content sheet, a packaging PDF and a few notes: scattered, partly incomplete, sometimes contradictory. They must turn this into one clear brief for a design agency. Done by hand, it is slow and error-prone.
What it does
| The problem statement asks for | How Brief AI handles it |
|---|
| Read the files provided | Parses .xlsx, .pdf and .docx into source elements (rows, pages) |
| Ask smart follow-up questions | A rule-based gap check produces a question only for items the extracted facts do not already answer |
| Identify inconsistent or unclear information | Rules look for conflicting facts across sources; the model also lists uncertainties it could not resolve |
| Create a brief an agency can use | A sectioned brief with confidence scores, plus open questions for the agency |
| Show where information came from | Every sentence carries citation chips (file + row or page); an Evidence panel shows the source text |
| Let the user review and edit before finalizing | Inline editing, a per-section review gate, then Finalize |
Architecture
- Ingest: files become atomic source elements (rows, pages).
- Extract: an LLM pulls facts, each with its source location.
- Rule checks (no LLM): gaps and conflicts are detected deterministically.
- Follow-up questions (no LLM): gaps are turned into questions, skipping anything the facts already cover.
- Draft: an LLM writes the brief with a confidence score per section.
- Self-critique: a second pass compares the draft with the extracted facts and lists discrepancies.
- Refine: the draft is corrected against those findings.
- Human review: evidence panel, inline edit, review gate, finalize.
Handling uncertainty and wrong AI output
- Each section shows a confidence score; sections with open findings are marked needs human review.
- The self-critique findings are shown under Uncertainty instead of being hidden.
- Gaps and conflicts come from rules, not from the model, so they do not depend on the model's mood.
- Every sentence is cited, so a reviewer can open the source and check it.
- Finalize stays locked until flagged sections are acknowledged.
- Known limit: the model can still add content the sources do not support (for example, design suggestions in our sample output), and the self-check does not catch everything. That is why a person signs off.
Models and technologies
| Role | Technology |
|---|
| Fact extraction, drafting, critique, refinement | Gemini (gemini-3.8-flash) |
| Fallback when quota is exhausted | Groq openai/gpt-oss-120b |
| Gap, conflict and follow-up logic | Plain Python rules (no model) |
| Backend | Python, FastAPI, Pydantic |
| Frontend | React, TanStack Start, Tailwind CSS |
| Hosting | Render (backend), Vercel (frontend) |
Run locally
Backend:
cd backend
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt
echo "GEMINI_API_KEY=your_key" > .env
echo "GEMINI_MODEL=gemini-3.8-flash" >> .env
python -m uvicorn main:app --reload --port 8000
Optional fallback: add GROQ_API_KEY and GROQ_MODEL=openai/gpt-oss-120b, and set LLM_PROVIDERS=gemini,groq.
Frontend:
cd frontend
npm install
echo "VITE_API_URL=http://127.0.0.1:8000" > .env.local
npm run dev
Open http://localhost:8080, click Use sample sources, then Generate brief.
Limitations
- Free API tiers. Daily and per-minute limits apply. Model responses are cached on disk, and [EDIT: keep only if you ship
sample_replay.py] for the bundled sample files the backend replays a saved output of a real pipeline run. Your own uploads run the live pipeline and may hit rate limits.
- Regenerate (after answering a follow-up question) needs live model calls and can fail when quota is used up.
- Text and tables only. Packaging artwork images are not read yet.
- Gap detection checks a fixed list of required fields (pack size, legal name, allergens, market, claims). It does not yet cover price or print run, and "already answered" is a keyword match.
- Sample data should not include confidential files, since free API tiers may use inputs to improve models.
Future work
- Read artwork images with a multimodal model.
- Regenerate only the affected section and keep user-edited sections locked.
- Export the finalized brief.
License
MIT
