I have been building AI applications to understand what happens when an LLM has to do more than just answer a prompt.
My projects cover research, agentic workflows, RAG, memory, automation, document processing, financial analysis, and AI-powered applications.
Here are some of the projects I have built and what each one taught me.
1. Deep Research Agent
Status: Live
GitHub: Research-AI-Agent
Live Demo: Research AI Agent
The Deep Research Agent takes a research topic and turns it into a structured research report.
Instead of making one LLM call, the application goes through multiple research stages. It generates search queries, collects information, extracts facts, identifies gaps, performs follow-up searches, and then uses the collected information to create the final report.
I built it with FastAPI and LangGraph, with Tavily for web research and ChromaDB for retrieval.
The project also supports report export, follow-up questions, summaries, counterarguments, translation, job history, Docker deployment, and an Ollama-based local mode.
What makes this project different:
The focus is on the complete research workflow rather than simply putting a chatbot interface on top of an LLM.
2. Autonomous Financial Research Agent
Status: Live
GitHub: Autonomous-Financial-Research-Agent
Live Demo: Autonomous Financial Research Agent
This project focuses on autonomous financial research.
The agent uses a ReAct-style workflow to decide what information it needs, select tools, process the results, and continue the research.
It can work with SEC EDGAR filings, earnings transcripts, financial data, news, sentiment, peer comparisons, fact checking, and calculations.
I also built three types of memory:
- Working memory for the current research session
- Semantic memory using FAISS
- Episodic memory for previous research runs
The project also includes conflict resolution, PII redaction, prompt-injection protection, rate limiting, and an evaluation framework.
What makes this project different:
It combines agent reasoning, multiple research tools, memory, financial data, and evaluation instead of treating research as a single search-and-answer task.
3. Autonomous Dental Appointment Bot
Status: Live
GitHub: Autonomous-Dental-Appointment-Bot
Live Demo: Autonomous Dental Appointment Bot
This project applies AI to appointment automation.
Patients can book, reschedule, and cancel appointments through web, SMS, WhatsApp, and voice interfaces.
The system connects the AI agent with application services, PostgreSQL, Redis, Celery, Stripe, and Google Calendar.
I also worked on practical problems such as appointment slot locking, payment webhooks, duplicate event handling, logging, health checks, and error handling.
What makes this project different:
The AI is connected to application workflows and business rules instead of functioning as an isolated conversation interface.
4. NexusBase
Status: Backend functional / Frontend being redeployed
GitHub: NexusBase
NexusBase is an enterprise RAG architecture project.
It uses Next.js, FastAPI, LangGraph, PostgreSQL, and pgvector.
The system routes queries through different paths depending on what the user is asking. Retrieval-based questions can go through vector search and response evaluation before producing the final response.
The entire application is containerized with Docker.
What makes this project different:
The project focuses on the architecture behind an enterprise RAG system, including routing, retrieval, evaluation, and structured responses.
5. MedComply
Status: Live
GitHub: MedComply
Live Demo: MedComply
MedComply is a medical compliance SaaS project.
It is structured as a monorepo with a Next.js frontend, FastAPI backend, and Supabase database migrations.
The system includes organizations, users, documents, authentication, role-based access control, document processing, and AI-assisted analysis.
The repository also includes Docker configuration and GitHub Actions for automated checks.
What makes this project different:
The project combines AI functionality with application-level access control, document management, and compliance-oriented workflows.
6. Aequitas FI
Status: Live
GitHub: Hybrid-Financial
Live Demo: Aequitas FI
Aequitas FI combines structured financial data with RAG.
Instead of treating every question as a document-search problem, the system separates structured SQL analysis from document retrieval.
LangGraph coordinates the workflow, while PostgreSQL and pgvector provide the data layer.
The project also includes temporal comparison, PII redaction, audit logging, human feedback, and automated testing.
What makes this project different:
It combines SQL and RAG in the same AI workflow so that structured data and retrieved documents can contribute to the final answer.
7. Context Synthesizer
Status: Architecture demonstration
GitHub: Context Synthesizer
Live Demo: Context Synthesizer
Context Synthesizer is an enterprise RAG architecture demonstration.
The project shows how information from systems such as Slack, Jira, Google Drive, and Notion could be organized into a unified retrieval workflow.
The interface demonstrates hybrid search, semantic retrieval, reranking, entity relationships, evaluation metrics, and observability.
This project is intentionally presented as an architecture demonstration.
The live external connectors, vector database, embedding pipeline, backend retrieval engine, authentication, and production LLM inference are not currently implemented.
What makes this project different:
The focus is on visualizing and explaining the architecture required for enterprise knowledge retrieval rather than claiming a demo interface is already a complete production system.
What I Learned From These Projects
The biggest thing I learned is that building an AI application is not just about connecting an LLM to a prompt.
The difficult parts are usually around the model.
You have to decide:
- What information the system can access
- Which tools it can use
- What it should remember
- How information should be retrieved
- How results should be checked
- What happens when something fails
- When the agent should stop
- How sensitive information should be handled
- How the system should be deployed
That is why my projects gradually moved from simple AI interfaces toward systems involving tools, retrieval, memory, evaluation, security, and application logic.
I am still improving these systems, but building them has given me practical experience with different ways of designing AI applications.
My goal is not just to make an LLM generate an answer.
I want to build systems around AI that can actually perform useful work.
My Projects
AI disclosure: This article was prepared with AI assistance for drafting and editing. The project information and technical details are based on my own projects and repositories, and I reviewed the content before publishing.
Top comments (0)