Data Scientist . AI & ML Engineer . Research Data Analyst
CIGMA Agentic Research Assistant
An end-to-end AI-powered research platform that combines Retrieval-Augmented Generation (RAG),
semantic document search, scientific literature retrieval, conversational memory, citation
generation, and automated report writing to help researchers discover, analyse, and synthesize
evidence from both private document libraries and live academic databases.
CIGMA Agentic Research Assistant is a production-ready AI research platform designed to
streamline scientific literature discovery, knowledge retrieval, evidence synthesis,
and technical report generation. Unlike conventional chatbots that rely solely on
pretrained knowledge, the system combines Retrieval-Augmented Generation (RAG),
semantic vector search, live scientific literature retrieval, and conversational memory
to generate evidence-based responses grounded in both private research collections and
trusted online academic repositories.
Researchers can upload their own collections of PDFs, DOCX documents, and text files to
build searchable research libraries. The platform automatically extracts text,
partitions documents into semantic chunks, generates vector embeddings using
Sentence Transformers, indexes them with FAISS, and retrieves the most relevant
passages whenever a question is asked. The retrieved evidence is combined with
current scientific publications from PubMed and OpenAlex before being synthesized by
a Large Language Model into structured, citation-aware reports.
Private research library creation with semantic document retrieval
Retrieval-Augmented Generation (RAG) for evidence-grounded responses
Real-time scientific literature retrieval from PubMed and OpenAlex
Automatic evidence ranking and query optimization
Conversational memory for multi-turn research discussions
AI-generated structured reports with references and supporting sources
Speech-to-text and text-to-speech capabilities for voice interaction
RESTful FastAPI backend with React frontend deployment
Core Capabilities
Upload and manage multiple research libraries containing PDF, DOCX, and TXT documents
Automatic document ingestion, chunking, embedding generation, and FAISS indexing
Semantic similarity search across uploaded research collections
Live retrieval of recent peer-reviewed literature from PubMed and OpenAlex APIs
Query rewriting and optimization to improve scientific search quality
Evidence ranking pipeline that prioritizes the most relevant publications
Citation extraction and automatic reference generation
Persistent conversational memory for context-aware research sessions
Professional AI-generated research summaries, analyses, limitations, and recommendations
Speech recognition using OpenAI Whisper and audio synthesis using Google Text-to-Speech
REST API supporting frontend integration and third-party applications
Designed for researchers, postgraduate students, clinicians, analysts, and R&D teams
Research Workflow
The platform follows a multi-stage agentic workflow that integrates local knowledge,
online scientific evidence, and conversational context into a single reasoning pipeline.
Every user query passes through retrieval, ranking, synthesis, and report generation
before a final evidence-based response is returned.
User uploads a personal research library consisting of scientific documents.
Documents are parsed, cleaned, chunked, embedded, and indexed using FAISS.
User submits a research question through the web interface or voice input.
The query is optimized to improve external scientific database retrieval.
Relevant passages are retrieved from the local vector database.
Recent publications are simultaneously retrieved from PubMed and OpenAlex.
Retrieved evidence is ranked according to semantic relevance.
The language model synthesizes local knowledge and external evidence into a structured report.
Conversation history is preserved to support follow-up questions and iterative research.
Final responses include supporting sources and automatically formatted references.
End-to-end workflow illustrating document ingestion, semantic retrieval,
scientific literature search, evidence ranking, LLM reasoning,
conversational memory, and structured report generation.
Architecture & API
The platform adopts a modular service-oriented architecture designed for scalability,
maintainability, and future integration with additional AI tools. The backend is built
with FastAPI and exposes RESTful endpoints that orchestrate document retrieval,
semantic search, scientific literature discovery, conversational memory, report
generation, and audio processing. Each responsibility is isolated into dedicated
services, enabling independent development and future extension.
Frontend built with React for an interactive research workspace and document management interface.
Library Service responsible for document ingestion, library creation, metadata management, and FAISS vector storage.
Retrieval-Augmented Generation (RAG) pipeline using Sentence Transformers embeddings and FAISS semantic search.
Research Agent orchestrating local knowledge retrieval, scientific web search, evidence ranking, conversational memory, and LLM reasoning.
Integrated scientific literature retrieval from PubMed and OpenAlex.
Conversation Memory Service maintaining context across multi-turn research sessions.
Whisper-powered speech transcription and Google Text-to-Speech for voice interaction.
Modular architecture enabling future integration of additional search providers, LLMs, and enterprise knowledge bases.
High-level architecture illustrating interactions between the React frontend,
FastAPI services, FAISS vector database, scientific search APIs,
Groq LLM, and conversational memory.
Modular REST API design supporting document management, intelligent
retrieval, research generation, export, audio processing, and health monitoring.
Core API endpoints include document library management, semantic retrieval,
intelligent agent execution, report export, speech transcription,
text-to-speech synthesis, and system health monitoring. The modular service
layer separates concerns across retrieval, memory, search, ranking,
citation generation, and language model inference, allowing each component
to evolve independently.
Deployment & Performance
Frontend deployed on Vercel for fast global content delivery.
Backend deployed using Render with FastAPI and Uvicorn.
Environment-driven configuration using secure environment variables.
Cross-Origin Resource Sharing (CORS) configured for secure frontend-backend communication.
Persistent vector libraries stored locally using FAISS indexes and metadata files.
Sentence Transformer embeddings initialized lazily to minimize application startup time.
Whisper speech recognition model loaded only when transcription is requested, significantly reducing server initialization latency.
Library processing modules imported lazily during document ingestion instead of application startup.
Service-oriented architecture optimized to reduce FastAPI cold-start time from approximately 18 seconds to under 2 seconds.
Scalable architecture prepared for Docker containerization, PostgreSQL integration, cloud object storage, and distributed vector databases.
Business Impact
Accelerates literature review by combining private research collections with live scientific publications.
Reduces manual evidence gathering through intelligent semantic retrieval and automated ranking.
Improves research reproducibility by grounding responses in verifiable sources and citations.
Supports postgraduate research, systematic reviews, grant preparation, and scientific writing.
Provides organizations with a reusable internal knowledge assistant for research and innovation teams.
Transforms disconnected documents into an intelligent, searchable organizational knowledge base.
Demonstrates production-grade AI engineering practices including RAG, vector databases, LLM orchestration, and service modularization.
