Research Buddy V1.1

Motivation

The Research Buddy project was born out of a personal need. After stepping away from research for some time, I recently resumed working in this space and immediately recognized the need for a system to simplify finding papers aligned with my interests. The starting point was an ML-based classifier for paper abstracts, giving me an immediate hint about a paper’s field before I invested time reading it in depth.

However, my requirements soon expanded. I realized that beyond classification, I also needed summarization and keyword extraction—tools that could provide quick insights into a paper’s content. While free tiers of established LLMs such as Perplexity, ChatGPT, or Gemini exist, they come with limitations: they are fragmented across platforms, and often behind paywalls for extended use. What I sought was an integrated system, built for me, all in one place, without additional cost.

This motivation pushed me to leverage my system design and full-stack development skills to create a modular full-stack ML/AI application.

Research Buddy v1.1 — Current Capabilities

The first official release (v1.1) of Research Buddy is capable of the following:

• Abstract Extraction
  Extract abstracts directly from uploaded PDF research papers.
• Paper Classification
  Predict the probable field of a paper using either a single ML/DL model or a comparative graph of multiple models with confidence level, providing transparency in predictions.
• Keyword Extraction
  Dual pipeline:
  Local Transformer model for keyword generation.
  Gemini LLM (via free API tiers) for high-quality keyword suggestions.
  Users can compare keywords side by side for better evaluation.
• Summarization
  Same dual pipeline approach: Transformer + Gemini LLM.
  Summaries displayed side by side for direct comparison.
• User Management & Libraries
  A robust user authentication and library system.
  Users can curate personal collections of papers.
  CRUD operations on papers: upload/save to CDN (e.g., AWS S3), store title/abstract, and attach preferred sets of keywords and summaries.

Case Study: Model Training and Dataset

For this project, I initially started with the entire arXiv dataset. However, due to the very limited computational capacity of my laptop and the restrictions of free-tier cloud resources, training quickly became impractical and nearly impossible. This resource bottleneck was a major pain point of the project.

To mitigate this, I created a Computer Science–specific subset of the arXiv data, focusing on five categories most closely aligned with my research interests:

• cs.CV – Computer Vision
• cs.CL – Computation and Language
• cs.IT / math.IT – Information Theory
• cs.LG – Machine Learning
• cs.RO – Robotics

This reduced dataset made it feasible to train multiple models and run comparative evaluations.

Model Comparison Evaluation

To illustrate, a randomly selected abstract from the dataset (Index 387, true category cs.CL) was classified across eight different models.

Abstract (excerpt):
Current task-oriented dialog (TOD) systems mostly manage structured knowledge (e.g. databases and tables) to guide goal-oriented conversations. However, they fall short of handling dialogs which also involve unstructured knowledge...

True Category: cs.CL

Model Predictions:

Observations

• 7 out of 8 models correctly classified the abstract as cs.CL.
• The only incorrect prediction came from XGBoost, which misclassified the paper as cs.RO.
• Feedforward NN and KNN gave the most confident correct predictions.
• BiLSTM performed comparably well despite requiring sequential tokenization.
• AdaBoost often predicted correctly but with low confidence, a recurring limitation.
• SVM classified correctly but lacks probability scores, limiting its use in confidence-based workflows.

Conclusion

• Traditional models (e.g., Random Forest, Multinomial NB) still perform reliably on structured TF-IDF inputs.
• Deep learning models (Feedforward NN, BiLSTM) deliver superior accuracy and confidence on clean dense representations or sequential inputs.
• XGBoost proved less suitable for this task, likely due to its inability to capture semantic and sequential relationships in natural language.

This evaluation demonstrates that even with a reduced dataset, the system effectively serves its intended purpose: providing quick, comparative classification results to guide paper exploration. However, to train a generalized, field-wide classifier, much larger resources and more comprehensive data would be necessary.

Design Principles

• Modularity: Both backend and frontend are structured for maintainability. Transformers or LLMs can be swapped out with minimal overhead.
• Scalability: Backend supports additional ML models and services.
• User-Centric: Designed first for personal productivity, but extensible to wider research communities.

Technology Stack

Research Buddy v1.1 is built on a modern, modular, and cost-conscious technology stack:

• FastAPI — Lightweight asynchronous Python backend framework, serving APIs for classification, summarization, keywords, and CRUD operations.
• React — Component-driven frontend with Material UI for a clean, responsive user interface.
• MongoDB — NoSQL database to store user accounts, libraries, paper metadata, abstracts, keywords, and summaries.
• CDN (AWS S3 + CloudFront) — Reliable storage and global distribution of uploaded PDFs and related assets.
• Gemini API — Integrated LLM for keyword and summary generation (leveraging free tiers; no vendor lock-in).
• Kaggle & Google Colab — Cloud-based GPU environments for training and experimentation with ML/DL models.
• MLflow — Framework for model lifecycle management (MLOps), enabling experiment tracking, versioning, and reproducibility across models.
• Local Transformers (Hugging Face) — Integrated for on-device keyword extraction and summarization, ensuring independence from external APIs and providing a fallback option alongside Gemini.

Rationale:
This stack balances low cost, flexibility, and scalability. Free-tier services (Gemini, Colab, Kaggle) lower experimentation costs, while MLflow and modular APIs ensure smooth upgrades when swapping or retraining models. AWS S3 + CloudFront provide reliable content delivery, while MongoDB ensures dynamic user data handling.

Future Development

• Conversational Features
  An open-source LLM is being fine-tuned for an integrated conversational feature. Users will be able to interact directly with a particular paper, with dialogue grounded in the document.
• Retrieval-Augmented Generation (RAG)
  The dialogue will leverage RAG, ensuring context-aware and document-specific responses.
• AI Agent Mode — Research Adviser
  A complete AI Agent mode is envisioned as a Research Adviser, unifying existing modules:
  • Classification
  • Summarization
  • Keyword extraction
  • Comparison
  Enhanced with upgraded features and proactive guidance.