InspirationFreelancers and independent consultants lose promising leads because responding to a new inquiry requires several context switches: understand the request, remember relevant past work, write a credible reply, check the calendar, and propose a call. FollowThrough turns that repetitive preparation into a reviewable draft while keeping the human responsible for every message.## What it doesFollowThrough is an AI agent for solo consultants, freelance designers and developers, and small service businesses. It:1. extracts the client's need, budget signals, urgency, and timezone hints;2. retrieves genuinely relevant past-project records from a portfolio store;3. uses Amazon Nova Lite on Amazon Bedrock to draft a concise proposal grounded in those records;4. suggests exactly three intro-call times; and5. stops at a visible human-approval gate.It never sends an email automatically. The reviewer can edit the proposal, request a revision, or approve it for the human to send.## How we built itThe workflow uses the AWS Strands Agents SDK in Python with Amazon Nova Lite (amazon.nova-lite-v1:0) through Amazon Bedrock in us-east-1. Deterministic tools parse the inquiry, search a local JSON portfolio, and provide demo-stable calendar slots. A lightweight Streamlit dashboard makes the extracted signals, matched work, editable draft, suggested times, and approval gate visible.The design intentionally separates repetitive preparation from professional judgment. If no portfolio match exists, the agent is told not to invent one. Calendar availability and email ingestion use controlled sample data for a reliable hackathon demo; they can later be replaced with Google Calendar freebusy and Gmail integrations.## Why the approval gate mattersA fabricated project reference, incorrect price, or wrong tone can damage a client relationship. FollowThrough prepares the work quickly but never auto-sends. Approval only marks the draft as ready for the freelancer to send themselves.## Challenges and learningsThe most important lesson was that an agentic workflow does not need to automate the final judgment call. Explicit orchestration, grounded retrieval, and a visible approval boundary make the system easier to inspect and safer to trust.## What's nextFuture work includes Gmail ingestion, Google Calendar availability, S3-backed portfolio history, follow-up draft suggestions, and deployment through Amazon Bedrock AgentCore Runtime.## Links- Source code: https://github.com/battinisadwik208-coder/followthrough

  • Demo note: the prepared demo uses controlled sample inputs and mock calendar slots; live Nova validation is pending AWS Bedrock account-access review.

Built With

  • amazon-bedrock
  • amazon-nova-lite
  • aws-strands-agents-sdk
  • json
  • python
  • streamlit
Share this project:

Updates

Submission history