Highlights

• Uses a production-style multi-agent workflow where a planner agent coordinates specialized agents for instrument tagging, report generation, chart creation, and retirement forecasting. The system splits analysis into focused tasks instead of relying on a single monolithic model call.
• Runs a research pipeline that continuously gathers market intelligence and stores it in a vector knowledge base for retrieval-augmented context.
• Deploys on AWS using serverless and managed services, including Lambda, App Runner, API Gateway, SQS, Aurora Serverless, SageMaker, and S3 Vectors.
• Provides an authenticated full-stack user experience with Clerk, FastAPI, and Next.js so users can manage accounts, positions, settings, and run asynchronous analysis jobs from the web app. Results appear as narrative reports plus interactive visualizations.

Highlights

• Built as an end-to-end autonomous agent loop that scrapes RSS deal sources, filters new opportunities against memory, and decides when to notify the user. The planning layer uses tool-calling with OpenAI to coordinate scanning, price estimation, and alerting.
• Utilizes a specialist pricing agent that was fine-tuned using QLoRA from an open-source LLM (llama3.2:3B) on an 800k dataset of product descriptions and prices.
• Uses an ensemble pricing strategy instead of a single model. A remote fine-tuned specialist model on Modal, a RAG-based frontier agent over ChromaDB, and a PyTorch neural regressor are combined to estimate fair product value more robustly.
• Implements retrieval-augmented pricing with a persistent vector database. Product embeddings stored in ChromaDB are queried for similar items, then injected into the pricing prompt as comparable market context.
• Includes an LLM preprocessing step to rewrite noisy product text into a cleaner structured description before pricing.

Highlights

• New inbound files can trigger the workflow automatically, starting data extraction and downstream AI actions. This turns raw inputs into operational workflows with minimal manual handling.
• A prospecting sub-agent finds qualified leads online and predicts likely contact emails. The results are then passed into CRM creation and outreach flows.
• Lead data is structured and pushed into Pipedrive, while an SDR sub-agent drafts outbound emails in Gmail. This creates a cleaner handoff from lead discovery to sales outreach.
• Additional sub-agents handle deal creation, demo booking, calendar checks, and notifications. This makes the system flexible for RevOps, sales, and business development use cases.

Highlights

• The system builds a persona from curated resume data, profile facts, summary notes, and communication-style inputs. This keeps responses aligned with the subject's background and tone rather than producing generic chatbot output.
• The implementation combines a Next.js frontend with a FastAPI backend that supports OpenAI and AWS Bedrock. Session memory and context injection help the assistant maintain continuity across user conversations.
• The app also functions as a lightweight lead-generation layer through visitor detail capture and unknown-question logging. That makes it useful as both a portfolio experience and a feedback channel for improving the twin over time.
• The project is production-oriented and deployed on AWS using Terraform-managed infrastructure. Delivery is automated through GitHub Actions, with the backend packaged for Lambda and the frontend served through the AWS static hosting stack.

Highlights

• Built as a graph-driven agent workflow with separate worker, tool, and evaluator stages, enabling multi-step reasoning, tool use, and self-correction in a controlled loop.
• Supports practical task execution through MCP servers for browser automation, sandboxed file management, Python REPL execution, web search, and Wikipedia retrieval, making it useful for research, lightweight automation, and assisted problem solving.
• Adds an automated quality check layer by scoring each response against user-provided success criteria, which is a strong pattern for reliable agent behavior and evaluative AI systems.
• Includes conversation persistence and session reset handling with in-memory checkpointing and resource cleanup, keeping the interactive experience stable during repeated use.
• Exposes the system through a simple Gradio chat interface, making the agent workflow accessible as a usable application rather than just a backend prototype.

Highlights

• Simulates four distinct AI traders with different investment styles, each running as an autonomous agent that can research, decide, and trade against its own portfolio.
• Uses MCP servers for account operations, market data, push notifications, web research, and memory, showing how tools and resources can be composed into a modular agent system.
• Persists balances, holdings, transactions, logs, and cached market snapshots in SQLite, giving each trader stateful behavior over time instead of one-off outputs.
• Includes tracing and live dashboarding with logs, portfolio value charts, holdings, and transaction history in Gradio.

Highlights

• Uses a Supabase vector store to fetch relevant product knowledge during conversations. This helps keep answers accurate and context-aware.
• n8n handles incoming questions, retrieval, and response delivery through a webhook-based workflow. This supports scalable product support and lead handling.
• Product data is pulled from Google Sheets, formatted, embedded, and stored in the knowledgebase. This makes updates simple and maintainable.
• Designed for conversational product support and qualification flows. It works well for sales, support, and inbound lead capture.

Highlights

• Built for healthcare consultation workflows, where doctors can enter visit notes and receive structured outputs for records, follow-up actions, and patient communication in one pass.
• Uses a streaming AI response flow so generated content appears progressively instead of waiting for a full completion.
• Implements user authentication and protected access with Clerk, including gated premium usage and embedded billing/pricing flows.
• Uses a separate FastAPI backend to validate authenticated requests and call the OpenAI API for note transformation.
• Structured for deployment across Vercel and AWS App Runner, combining a modern frontend delivery model with a containerized Python service.

Highlights

• Supports a catalog-driven legal drafting flow across 12 agreement types, including NDAs, cloud service agreements, pilot agreements, data processing agreements, and AI addenda. The system guides users toward the closest supported template when an exact request is not available.
• Uses structured LLM outputs to separate document selection from draft generation, making the workflow safer and more controllable than freeform chat. The backend also includes provider-routing fallback logic to keep the drafting flow resilient across model/API differences.
• Combines chat, live markdown preview, autosave, saved sessions, and PDF export in one workspace.
• Runs as a full-stack containerized app with FastAPI serving a statically built Next.js frontend and a reset-on-start SQLite database.