Tailor Mail

A Chainlit chat app that turns your resume and a job posting into a tailored cold email or cover letter — drafted in your voice, not an AI's.

How it works

1. Onboarding — paste your resume (text or PDF). It's parsed into a structured profile (skills, experience, education, social links) and enriched with your GitHub bio and top repos. Saved locally so you only do this once.
2. Apply — paste a job posting URL or description. It extracts the company, role, requirements, and posting tone, then streams a drafted email or cover letter live in the chat.
3. Refine — say things like "more casual", "more formal", "simpler vocab", or "richer vocabulary" to redraft in a different tone or word choice, reusing the same extracted job facts (no re-scraping).

Tone is asked once via action buttons and remembered for next time (data/preferences.json); say a tone name again to change it.

Setup

(Windows / PowerShell)

.venv\Scripts\Activate.ps1
pip install -r requirements.txt

Copy .env.example to .env and fill in OPENAI_API_KEY (and GROQ_API_KEY if/when the Groq branch in utils/llm.py is wired up).

Run

chainlit run app.py -w

-w enables auto-reload on file changes.

Project layout

• app.py — Chainlit handlers (on_chat_start / on_message) and the session state machine (resume onboarding, job detection, tone/vocab follow-ups)
• graphs/ — compiled LangGraph flows: profile_graph (resume → structured profile → GitHub enrichment → save) and apply_graph (job input → fetch → structured job facts)
• nodes/ — graph nodes: extract.py (structured Profile/JobFacts via LLM), scrape.py (job fetching + GitHub enrichment), draft.py (streaming email/cover-letter generation)
• prompts/ — system.py (voice/tone fragments, list of AI-sounding patterns to avoid) and templates.py (composed system prompts) — tweak these to change output style
• utils/ — llm.py (single LLM provider factory — see get_llm), scraper.py (httpx + BeautifulSoup page fetching, PDF text extraction via pypdf, GitHub API lookups)
• memory/ — flat-JSON persistence layer for the user's profile and preferences (data/profile.json, data/preferences.json; gitignored)

For a deeper architectural walkthrough (state machine details, LLM task routing, prompt composition), see CLAUDE.md.

Notes

• Only OpenAI is wired up by default (gpt-4o-mini for extraction, gpt-4o for drafting); a Groq branch is sketched in utils/llm.py.
• LinkedIn profile enrichment is intentionally skipped (LinkedIn blocks scraping) — that data comes from your resume text only.
• There is no test suite, lint config, or CI in this repo.