Internship posts checked by AI and alerts sent on WhatsApp

Internship posts checked by AI and alerts sent on WhatsApp

An n8n workflow for students who follow internship channels on Telegram. You forward a post to your bot. An AI model splits it into separate openings, the workflow reads each application page, and a second AI model decides whether the opening fits pre-final year students of the batch of 2027. A fitting one is logged, set as a calendar reminder for its deadline and sent as a WhatsApp message to the numbers on your list. Thirty nodes do the work.

What is inside

• workflow.json: the workflow as n8n exports it, 30 nodes and no sticky notes. It holds no credential.
• README.md: these notes.

What we changed from the original copy

Mindket repaired and cleaned the copy:

• Three addresses pointed at the previous owner's own services: a Firecrawl server, a webhook that wrapped a search, and a WhatsApp sender. The Firecrawl step now calls https://api.firecrawl.dev/v2/scrape with a Header Auth credential. The search tool now calls SerpApi directly and asks it for only the title and snippet of each result. The WhatsApp address is the placeholder <YOUR_WHATSAPP_SEND_URL>.
• The previous owner's Google Sheet, data table and calendar ids are now <YOUR_GOOGLE_SHEET_ID>, <YOUR_DATA_TABLE_ID> and <YOUR_GOOGLE_CALENDAR_ID>. The three sheet tabs are picked by name: Mobile Numbers, True Classification and False Classification.
• The Groq model llama-3.3-70b-versatile was shut down on 16 August 2026. Groq's deprecations page, read on 12 November 2026, names openai/gpt-oss-120b as the replacement, so Mindket's copy uses it. The Groq model openai/gpt-oss-20b stays.
• The Cerebras models llama3.1-8b and qwen-3-235b-a22b-instruct-2507 were deprecated on 27 May 2026. Cerebras names gpt-oss-120b as the replacement, and its model list shows it, so both Cerebras nodes now use it.
• The Notify User node ran a workflow named Agent Error from the previous owner's n8n. Its id is cleared, so pick a workflow of your own or delete the node.
• A stray column name, Internship, is gone from the data table's matching columns.
• The file had no settings. Mindket's copy uses the current execution order, v1.
• No credential is inside.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• A Telegram bot made with @BotFather.
• A Groq API key, a Cerebras API key, a Firecrawl API key and a SerpApi API key.
• A Google Sheet with three tabs. Mobile Numbers has a column Phone Number, filled with digits that start with the country code and no plus sign. True Classification has Timestamp, Telegram Message, Title, Company Name, Eligibility, Application Link and Status. False Classification has Timestamp, Telegram Message, Application Link and Rejection Reason.
• An n8n data table with the columns title, Company_Name, Eligibilty and Application_Link. The third name is spelled that way in the file.
• A Google Calendar, and a Google account for the credentials.
• A WhatsApp sender of your own that accepts a POST with the fields to and message.

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 Resend failed message.
3. Add a Groq credential and select it in Splitter Main and Analyst Fallback.
4. Add an OpenAI credential whose base URL is https://api.cerebras.ai/v1 and whose key is your Cerebras key. Select it in Splitter Fallback and Analyst Main.
5. Add a Header Auth credential for Firecrawl: the name Authorization and the value Bearer followed by a space and your key. Select it in Firecrawl Scrape.
6. Add a SerpApi credential and select it in serpapi_request.
7. Add a Google Sheets credential. In Get Mobile Numbers, Append Valid and Append Rejected, replace <YOUR_GOOGLE_SHEET_ID> with the id of your sheet.
8. Create the data table. In check_duplicate, Check for duplicate valid and Insert row, replace <YOUR_DATA_TABLE_ID> with its id.
9. Add a Google Calendar credential. In Add event reminder, replace <YOUR_GOOGLE_CALENDAR_ID> with your calendar's id.
10. In Send Whatsapp Message, replace <YOUR_WHATSAPP_SEND_URL> with the address of your sender.
11. In Notify User, pick a workflow of your own, or delete the node.
12. Change the year 2027 and the role rules in the prompt of Analyst Agent, in Pre-Check and in the name of For 2027?, if your batch is another.
13. Publish the workflow, forward one internship post to your bot and check each tab and the data table.

How a run goes

1. Telegram Trigger starts a run for every message the bot receives.
2. Text Guard passes only a message whose text is longer than 20 characters.
3. Splitter Agent asks Splitter Main to split the text into a JSON list, one object for each opening, with role, company, stipend, eligibility, description and apply_link. It skips channel invitation links. It tries up to 5 times, and Splitter Fallback takes over if the main model fails. On an error, Notify User runs.
4. JSON Format Fixer removes a Markdown fence and parses the list into one item for each opening.
5. Loop Over Each Link takes one opening at a time.
6. Firecrawl Scrape fetches the application page as Markdown. If the site blocks it, the run goes on with an empty result.
7. Pre-Check keeps the first 30,000 characters of the page. When the page is shorter than 150 characters, it uses the Telegram text instead. It rejects an opening whose text names a year from 2020 to 2026 and not 2027. For 2027? sends the rest to the analyst and the rejected ones to Merge.
8. Analyst Agent decides, with Analyst Main and Analyst Fallback. Its prompt asks it to call check_duplicate first. It calls serpapi_request only when the batch year is missing. It answers with JSON: send_whatsapp, reason, whatsapp_message and deadline. On an error, Wait pauses for 2 minutes and Resend failed message sends the post's text back to you.
9. Output Parser pulls the JSON out of the reply. Check for duplicate valid looks up the link in the data table, and Not Duplicate? passes the opening only when send_whatsapp is true and no row has that link.
10. Deadline Exists? sends an opening with a deadline to Add event reminder, which creates a calendar event from 09:00 to 10:00 UTC on that day, titled with APPLY TODAY.
11. Get Mobile Numbers reads the numbers, and Loop Over Items sends the message to each one through Send Whatsapp Message.
12. Append Valid logs a row with the status Pending, and Insert row stores the opening in the data table. A rejected or repeated opening goes to Append Rejected with its reason. Then the loop moves to the next opening.

Change the look

• The kind of internship: the prompt of Analyst Agent, the Pre-Check code and the name of For 2027?.
• The text of the alert: the template in the prompt of Analyst Agent.
• The channel: swap Send Whatsapp Message for a Telegram or email node.
• The deadline reminder: the times in Add event reminder, or delete the node.
• The models: the four chat model nodes.

Good to know

• The workflow is built for one reader, a pre-final year student of the batch of 2027 who wants technical internships. Change the prompts and Pre-Check for anyone else.
• The Cerebras nodes are OpenAI Chat Model nodes pointed at Cerebras through the credential's base URL. Cerebras's own page, read on 12 November 2026, gives https://api.cerebras.ai/v1 for that.
• WhatsApp's Business Messaging Policy, read on 12 November 2026, says you may contact people only when they gave you their number and agreed to hear from you. Keep the Mobile Numbers tab to people who asked, and treat the numbers as personal data.
• The text of each post goes to Groq and Cerebras, the application link goes to Firecrawl, and a search phrase may go to SerpApi. Do not forward private messages to the bot.
• Read the terms of each application site before you scrape it. A site that blocks the request leaves the Telegram text as the only input.
• Each opening costs up to two model calls with retries, one Firecrawl request and sometimes one SerpApi search.
• Mindket did not run the workflow, the Firecrawl v2 call or the SerpApi call. It read the vendors' pages and n8n's source instead, so check the first runs by hand.
• The JSON holds no credential.