AI Enterprise SolutionsGPT & Claude

Orion

Students preparing for exams and organizations onboarding/training personnel struggle with information overload from hundreds of pages across unorganized PDFs, lecture slides, and official documents. Finding topic-specific core concepts and relevant practice questions requires hours of manual searching through redundant filler text, leading to inefficient learning and poor exam readiness. Existing AI models often hallucinate or provide generic answers outside the provided study materials, making them untrustworthy for strict academic or internal policy compliance.

GitHub

Preliminary score

54.5/ 100Earlyaverage of 2 judges

Value for the user14.5 / 25
A specific problem and a clear outcome from solving it.
Prototype and use of AI19.5 / 30
A working core scenario, and what AI actually contributes to the result.
Quality testing9.5 / 20
Tests, examples of failures, and comparison with the current approach.
Feasibility7.0 / 15
Data requirements, running costs and a clear next step.
Originality4.0 / 10
A meaningful difference from common solutions.

What the judges wrote

Claude Opus 5.5Preliminary49/100
Strengths
Your README describes a credible end-to-end RAG pipeline: PyMuPDF extraction with OCR fallback (including Azerbaijani), noise stripping, sentence-aware chunking, pgvector retrieval via an RPC, and schema-validated JSON output. The anti-hallucination guardrails are concrete engineering rather than slogans. The LLM is skipped when retrieval is empty, citations to non-existent passages are removed, and uncited concepts are dropped. You also honestly reported real failures, such as multi-column PDF extraction noise and out-of-scope queries, and explained how you handled them.
To improve
Your testing needs numbers. Build a small labelled set of topic queries over a few real PDFs, including out-of-scope ones, and report retrieval hit rate, citation accuracy and refusal correctness. Time a person doing the same task manually so you can back up the claim of 'hours to seconds'. You entered the enterprise track but never named one company workflow or the adoption metric the brief asks for, so pick a concrete case such as onboarding policy Q&A and quantify it. Add running costs per document and per query, and make your model claims consistent: the submission says Llama 3.1 70B, while the README defaults to gpt-oss-120b.
GPT-6 AstraPreliminary60/100
Strengths
Your PDF-to-study-pack workflow addresses a concrete information-search problem, with AI contributing directly through retrieval and structured generation. Your README specifies useful safeguards: page citations, schema validation, citation filtering, and a no-evidence path that skips generation. Your testing write-up identifies concrete failures in database setup, complex PDF layouts, and out-of-scope queries rather than claiming universal success.
To improve
Provide reproducible input/output examples showing ingestion, source-supported study questions, and an appropriate refusal when evidence is missing. Evaluate a labeled query set and report retrieval accuracy, citation correctness, unsupported answers, and time saved versus manual searching. For the enterprise track, select one specific onboarding or training workflow and validate its adoption metric with intended users. Add estimated ingestion and query costs, document-access and retention controls, and a concrete next pilot step.

The team

Ismail Alhumbatov
developer
Fakhraddin Abbasov
product
Farid GojamanliLead
developer

Disclosure

Required by the rules. Judges verify it against the repository.

Models: Llama 3.1 70B (via Groq API), Sentence-Transformers (all-MiniLM-L6-v2)
Data: Public educational PDFs, academic lecture notes, sample exam question banks, and synthetic test datasets.
Components: FastAPI, Supabase (PostgreSQL + pgvector), PyMuPDF (fitz), Groq API, Sentence-Transformers, Pydantic
Built during the event: Confirmed

Quality testing

We tested end-to-end PDF ingestion, semantic vector retrieval via Supabase pgvector, and grounded study material generation using Groq (Llama-3.1). What worked: 1. Fast parsing and vector storage for standard multi-page lecture PDFs. 2. High retrieval accuracy for direct topic queries, generating concise theory and practice questions strictly within document bounds. What failed & edge cases found: 1. Database Schema Dependency: During initial testing, vector chunk insertion failed (404) because the 'document_chunks' table hadn't been initialized via schema.sql in Supabase. 2. Complex PDF Layouts: PDFs with multi-column layouts and inline diagrams suffered from text extraction noise, requiring stricter chunk filtering. 3. Out-of-Bound Queries: Queries asking about topics omitted from the uploaded PDFs initially yielded empty context, which we handled by adding fallback prompt guardrails to state "Insufficient information in uploaded files" instead of hallucinating. Comparison with traditional methods: Currently, students manually skim 100+ pages across unorganized PDFs, taking hours to find core topics. Our solution automates noise reduction and topic-based extraction into seconds.

Worth 20 of the 100 points.

Project ownership

The team keeps the rights to this project. Judges were given access for verification only.