Engineered, Developed, and Designed by Neksha DeSilva
Licensed under Apache-2.0 Contributor's License
CatalystAI 2025 is a cutting-edge AI-assisted academic reasoning framework, purpose-built to analyze, interpret, and generate deep, context-aware engagement with uploaded documents. Leveraging the power of Gemini Pro, LangChain, and AstraDB, it operates as an intelligent assistant, educator, and evaluator — delivering comprehensive academic analysis through both textual and voice-driven interactions.
This framework is not a chatbot. It is an autonomous academic engine capable of cross-referencing documents, scoring understanding, generating expert-level questioning, and producing structured session archives for revision, research, and model training.
#Demo
Features: Query oriented answers, Document Completiton Score
Features: Accuracy Score
Features: Deep search feature, Expert Mode
Features: Web search, Document search override
Features: Web search, JSON Exporting, Chat Exporting ability
CatalystAI performs real-time tracking of how much of the uploaded document has been meaningfully addressed.
Powered by:
- Full-text indexing and tokenization of PDF content.
- Semantic overlap detection between AI-user conversation and original text.
- A refined completion logic model which tracks both coverage and contextual relevance.
The result is a Completion Score (in %) representing actual engagement with the source material — not just surface-level mentions.
CatalystAI uses large language model alignment algorithms to evaluate the fidelity of each AI-generated answer.
The Accuracy Score is derived from:
- Vector-based similarity between the generated answer and the document source.
- Cross-section validation across multiple portions of the PDF.
- Gemini Pro’s semantic matching engine for higher-order reasoning validation.
Additionally, the Deep Search Mode dynamically activates when context requires wide-range understanding across chapters, sections, or disconnected mentions, creating an intelligent context window before responding.
Expert Mode transforms CatalystAI into a self-interrogating academic entity:
- Automatically generates a bank of intelligent, context-relevant questions from the document.
- Immediately provides model answers to those questions.
- Simulates expert-level comprehension and tutoring behaviors.
Use cases include exam preparation, rapid revision, research validation, and peer simulation. This mode supports multiple levels of depth, including surface, interpretive, and critical evaluation questioning.
CatalystAI offers two primary working paradigms:
Power Assistant-GPT Mode
General-purpose reasoning and interaction using Gemini Pro — excels in synthesis, summarization, and freeform academic support.
Document Answering Mode
Bound to uploaded PDF context only. All responses are strictly generated with respect to the document, favoring integrity and precision over speculation.
Mode toggling is instant and session-aware.
CatalystAI includes full voice interaction capability, allowing:
- Voice-to-text conversion via microphone.
- Natural language understanding from spoken prompts.
- Spoken question parsing and intelligent document referencing.
Ideal for accessibility, mobile learning, and users seeking a more fluid input experience.
Powered by Web Speech API with domain-specific tuning.
Sessions are exportable as clean, structured .json files containing:
- User questions
- AI responses
- Expert mode questions + answers
- Completion and accuracy scores
- Session metadata (timestamps, document title, document hashes)
These exports are designed for:
- Academic review
- Research traceability
- LLM fine-tuning
- Institutional training feedback
Each session is saved in structured format with versioning support. Benefits include:
- Full playback of academic thought process
- Time-based learning progress tracking
- Re-assessment from any conversation point
- Cross-session performance analytics (planned)
Chat history is stored in both human-readable markdown and machine-parseable JSON formats for flexibility.
A postgraduate student uploads a 6000-word dissertation on neural cryptography.
CatalystAI:
- Performs full document parsing and token indexing.
- Calculates a Document Completion Score of 82% after several exchanges.
- Measures an Accuracy Score of 94% based on aligned source content.
- Activates Expert Mode, auto-generating 12 conceptual questions from the methodology section, and providing answers with references to exact document lines.
- The session is exported as a
.jsonfile for supervisor review and future fine-tuning.
| Component | Purpose |
|---|---|
| Gemini Pro (via LangChain) | LLM for NLU, semantic validation, and generative tasks |
| LangChain PromptTemplate | Structured prompt building and contextual memory handling |
| AstraDB | Vectorized document storage, indexing, and metadata control |
| JavaScript + HTML/CSS | Frontend interaction, interface logic, voice input handling |
| JSON Exporter | Clean archival and machine-readable export generation |
| Web Speech API | Voice input integration (transcription and intent parsing) |
CatalystAI will evolve into a research-centric co-pilot with the following capabilities:
- Local LLM fallback for offline inference and air-gapped academic use.
- Citation Mode that attaches exact page + line number to AI outputs.
- Collaborative Learning Mode allowing multiple students to engage in a co-session with synchronized threads.
- Domain Plugins such as equation solvers, code review interpreters, and multilingual translators.
- Dynamic Curriculum Engine that builds personalized learning goals from documents.
Apache-2.0 — Contributor’s License Edition
You are free to fork, build, and deploy CatalystAI for non-commercial or research purposes. For redistribution or monetization, a contributor citation and license retention is required.
Neksha DeSilva
LinkedIn Profile | CEO, Founder @ Indexx Inc.
Colombo, Sri Lanka
CatalystAI is not just a document assistant.
It is a scalable academic intelligence engine — capable of learning, reasoning, and teaching.