added portfolio description and updated readme

Author: PierreExeter

README.md

@@ -27,24 +27,24 @@ flowchart TD
     classDef db fill:#dc2626,stroke:#b91c1c,color:#ffffff
     classDef text fill:#ffffff,stroke:#4b5563,color:#1f2937
 
-    UI["Streamlit UI"] --> Agent["LangChain Agent"]
+    UI["Streamlit UI"] --> AGENT["LangChain Agent"]
     
-    subgraph UI
+    subgraph UI["Streamlit UI"]
         A[Inbox view<br/>Draft review<br/>Calendar view<br/>Settings]
     end
 
-    subgraph Agent
+    subgraph AGENT["LangChain Agent"]
         B[Email classification<br/>Email drafting<br/>Calendar scheduling]
     end
 
-    Agent --> gmail["Gmail API"]
-    Agent --> calendar["Calendar API"]
-    Agent --> hf["HuggingFace LLM API"]
-    Agent <--> db[(SQLite DB)]
+    AGENT --> gmail["Gmail API"]
+    AGENT --> calendar["Calendar API"]
+    AGENT --> hf["HuggingFace LLM API"]
+    AGENT <--> db[(SQLite DB)]
     
     class UI ui
     class A text
-    class Agent agent
+    class AGENT agent
     class B text
     class gmail,calendar,hf api
     class db db

docs/portfolio.md

@@ -0,0 +1,25 @@
+# Gmail Agent — AI-Powered Email & Calendar Assistant
+
+## Situation
+
+Managing a busy inbox involves repetitive tasks — classifying emails, drafting replies, and scheduling meetings — that consume significant time but follow predictable patterns. Modern LLMs have made it possible to automate these workflows intelligently, but most solutions operate as black boxes without meaningful human oversight.
+
+## Task
+
+I built a full-stack AI assistant that automates email classification, reply drafting, and meeting scheduling while keeping the user firmly in control. The project served as a deep dive into LLM application development, integrating LangChain orchestration, Google APIs, and a human-in-the-loop approval system into a cohesive product.
+
+## Action
+
+Built a **Streamlit** web app with a modular architecture:
+
+- **AI Agents** (LangChain + HuggingFace Llama 3.1): An email classifier that categorises messages (needs reply, FYI, meeting request, action item) with confidence scoring; a reply drafter with tone control and iterative improvement; and a meeting scheduler that extracts details and proposes free calendar slots.
+- **Human-in-the-Loop Approval**: Every action (sending emails, creating calendar events) requires explicit user approval. A risk assessment layer automatically flags unknown senders, sensitive content, and low-confidence classifications.
+- **Google API Integration**: Full OAuth 2.0 flow with Gmail (fetch, send, label) and Google Calendar (availability checking, event creation, free slot detection with working-hours awareness).
+- **Persistence & Learning**: SQLAlchemy/SQLite database tracking classifications, drafts, approvals, and user feedback to build an audit trail and support future model improvement.
+- **Graceful Degradation**: Every AI agent includes fallback strategies (heuristic keyword matching, template-based drafts) so the app remains functional even when the LLM is unavailable.
+
+## Result
+
+A fully functional, production-structured application that demonstrates end-to-end LLM integration — from prompt engineering and chain composition to API orchestration and state management. The security-first design (layered approvals, sensitive content detection, trusted sender lists) reflects real-world considerations for AI-assisted communication. The project solidified practical skills in **LangChain**, **Google Cloud APIs**, **OAuth 2.0**, **SQLAlchemy**, and **Streamlit**, with clean separation of concerns across agents, services, database, and UI layers.
+
+**Tech stack:** Python · LangChain · HuggingFace · Streamlit · SQLAlchemy · Gmail API · Google Calendar API · OAuth 2.0