Freelancers and independent consultants lose countless hours every week doing administrative busywork: reading vague project inquiries, calculating if a client's budget makes sense, and drafting repetitive emails to either accept, reject, or ask for more details. We wanted to build a tool that acts as a ruthlessly efficient digital assistant, allowing freelancers to focus on the work they actually get paid to do.
What it does FeatherTriage is an end-to-end lead evaluation and response dashboard:
Smart Ingestion: It connects directly to your Gmail and pulls in unread project inquiries.
Personalized Scoring: It evaluates the client's email against your specific profile settings (Ideal Wage, Ideal Working Hours, and Tech Stack).
AI Triage: It scores the lead on Budget Fit, Scope Clarity, Project Fit, and Scam Likelihood, then categorizes the email into High Match, Clarification Needed, or Low Match/Scam.
Auto-Drafted Replies: It automatically writes a tailored response based on its evaluation. If the lead is great, it drafts an acceptance email with a timeline. If it's vague, it drafts an email asking specific clarifying questions.
One-Click Execution: From the Streamlit dashboard, you can review the scores, tweak the AI's drafted email, and hit "Send" instantly.
How we built it We built the backend logic in Python, utilizing the simplegmail library to safely ingest unread messages and send outbound replies.
For the core intelligence, we utilized Featherless.ai. By hooking into their OpenAI-compatible API, we powered our triage engine with advanced open-source models (like Llama 3). We combined Featherless with the instructor library to guarantee perfectly structured JSON outputs. This was critical—our system needed the LLM to return strict, predictable data types (like integers for budget scores and booleans for timeline feasibility) so our frontend could render the data without crashing.
The frontend dashboard was rapidly prototyped and deployed using Streamlit, providing a clean, color-coded interface with interactive expanders and text areas for seamless email editing.
Challenges we ran into The biggest challenge was ensuring the AI didn't hallucinate structural formats. Early on, passing raw text directly to the UI caused errors if the LLM forgot a comma or changed a variable name. Switching our backend to enforce Pydantic schemas via instructor and routing it through Featherless's fast inference API completely solved this, giving us 100% reliable JSON payloads every single time.
Accomplishments that we're proud of We successfully created a completely closed-loop system. The app handles the entire lifecycle of a cold lead: fetching the raw data, applying complex reasoning based on user-defined constraints, sorting the data intelligently, generating a context-aware artifact (the reply), and executing the final send.
What we learned We learned how to effectively chain LLM prompts. We split our AI operations into two distinct steps: one focused entirely on analytical scoring (the Lead Analyzer) and a separate prompt focused entirely on creative writing (the Auto-Responder). This separation of concerns made the outputs drastically better.
What's next for FeatherTriage We want to add a calendar integration so the AI can look at the freelancer's actual schedule to automatically determine if a project's timeline is feasible before drafting the acceptance email.
Built With
- featherless-ai
- instructor
- meta-llama-3
- python
- simplegmail
- streamlit
Log in or sign up for Devpost to join the conversation.