Telegram resume and cover letter bot with AI and LaTeX

Telegram resume and cover letter bot with AI and LaTeX

An n8n workflow that turns a Telegram bot into a resume and cover letter assistant. A person sends a resume as a PDF, or just types details about themselves. An AI model reads it and keeps a profile for that person in Supabase. Commands then show the profile, tailor it to a job description, check it against the job's keywords, write a cover letter and send back compiled PDF files. Ninety-eight nodes do the work.

What is inside

• workflow.json: the workflow as n8n exports it, 98 nodes and no sticky notes. It holds no credential.
• README.md: these notes.

What we changed from the original copy

Mindket cleaned the copy and swapped the retired models:

• The seven main chat models used qwen3-coder:480b and rnj-1:8b on Ollama's cloud. Ollama's cloud model list, read on 12 November 2026, shows neither. The four larger ones now use gpt-oss:120b and the three smaller ones use gpt-oss:20b, both on that list.
• The seven fallback models used the Cerebras model zai-glm-4.7, which Cerebras deprecated on 17 August 2026. They now use gpt-oss-120b, which is on Cerebras's current shared list.
• Every credential is removed. The Telegram, Supabase, SSH, Ollama and Cerebras credentials are yours to add.
• The file had no settings. Mindket's copy uses the current execution order, v1.
• No host name, address or key was in the file, so there was nothing else to replace.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• A Telegram bot made with @BotFather.
• A Supabase project with a table named Profiles and four columns: telegram_id (a bigint), telegram_username (text), master_profile (text) and tailored_profile (text).
• An Ollama account with an API key, to call its cloud models.
• A Cerebras API key, for the fallback models.
• A Linux server you can reach over SSH with a password, with a full TeX Live install. The template uses packages such as cfr-lm, marvosym and tcolorbox, and the commands pdflatex and base64 must work there.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Add a Telegram credential with your bot token. Select it in Telegram Trigger and in every Telegram node. The nodes that say they need one are marked in n8n.
3. In Supabase, create the Profiles table described above. Add a Supabase credential with the project URL and the service role key, and select it in the seven Supabase nodes. Keep the table out of public access, because the service role key skips row security.
4. Add an OpenAI credential whose base URL is https://ollama.com/v1 and whose key is your Ollama API key. Select it in Ollama, Ollama JD, Ollama Fix LaTeX Model, Ollama CL Model, Classifier Model, Classifier Model1 and Ollama ATS Model.
5. Add a second OpenAI credential whose base URL is https://api.cerebras.ai/v1 and whose key is your Cerebras key. Select it in the seven nodes whose names start with Cerebras.
6. Add an SSH credential for your server, with its host, port, user name and password. Select it in Write LaTeX File, Write LaTeX File (Retry), Read PDF Base64 and Compile Cover Letter.
7. Publish the workflow so the Telegram trigger stays on.
8. Open your bot in Telegram, send /start, then send a resume as a PDF.
9. Send /view, then /download, and check that a PDF comes back.

How a run goes

1. Telegram Trigger takes every message and every button press. It downloads any file that comes with it.
2. Supabase Fetch by ID and Supabase Fetch by Username look for the person's profile row. Check Migration Needed gives a row found only by user name the person's numeric id.
3. Assess Master Profile checks that the profile has a name, a contact, a summary, skills, education, work experience and projects. Until it does, most commands answer with what is missing.
4. Command Router sends the message down one of sixteen branches. They cover a greeting, /start and /info, an incomplete profile, media the bot cannot read (a photo, sticker or voice note), /view, /clear, a PDF upload, chat, /download, /downloadoptimized, /optimize, /viewoptimized, /coverletter and /analyze. Two more branches warn when /optimize or /analyze comes with no job description.
5. A PDF upload goes through Extract PDF Text and AI Agent. The agent fills in the profile and Parse JSON, Safe Profile Merge and Row Exists? save it with Supabase Update or Supabase Create. A file that is not a readable PDF gets an error message.
6. Free text goes to Intent Classifier, and Parse Intent also checks the first words of the message. A question goes to Query Agent, which answers only from the saved profile. Anything else goes to AI Agent, which updates the profile, with Window Buffer Memory keeping the last turns of that chat.
7. /optimize followed by a job description goes to AI JD Agent. It writes a tailored copy of the profile and Supabase Save Tailored stores it. The master profile stays as it was.
8. /analyze followed by a job description goes to ATS Gap Agent, which lists the keywords the profile lacks. The reply is split to fit Telegram's message limit.
9. /download and /downloadoptimized run Generate LaTeX, which fills a resume template, and Write LaTeX File sends it to your server. The server compiles it twice with pdflatex. On success Read PDF Base64 returns the PDF, and Send PDF (Telegram) and Send LaTeX Source (Telegram) deliver both files.
10. When the compile fails, AI Fix LaTeX Agent rewrites the source, Write LaTeX File (Retry) tries once more, and Send Compile Error Msg tells the person if that fails too.
11. /coverletter uses AI Cover Letter Agent, which writes the letter from the tailored profile and its job description, then Compile Cover Letter builds the PDF the same way.
12. /clear runs Supabase Clear Profile and deletes the row at once.

Change the look

• The resume layout: the template in Generate LaTeX.
• The cover letter layout: the template in Generate LaTeX Cover Letter.
• The voice of the cover letter: the system message of AI Cover Letter Agent.
• What counts as a complete profile: the checks in Assess Master Profile.
• The welcome text and the command list: Telegram Info Reply.
• The models: the seven chat model nodes and their seven fallbacks.

Good to know

• Resumes are personal data. Each one is stored in your Supabase table, and its text goes to Ollama and to Cerebras. Tell your users, and delete rows when someone asks.
• /clear has no confirmation step. A press on it deletes the profile.
• The SSH nodes run shell commands on your server. The LaTeX text is passed as base64, so what the model writes is not run as a command. Use a server and a user account that hold nothing else.
• Files are written to /tmp with the execution number in the name and removed after a successful run. A failed run can leave files behind.
• The Telegram messages use HTML formatting, and the workflow escapes the model's text before sending it.
• Each AI node retries up to 3 times and then falls back to its Cerebras model.
• Mindket did not run the workflow, the Ollama and Cerebras calls, or the SSH step. It read n8n's source and the vendors' pages instead, so test every command once by hand.
• The JSON holds no credential.