A live WhatsApp bot, and the tool that tailored my CV

The problem

What I did

• Built and launched a Malayalam WhatsApp bot for an Akshaya e-Centre on Meta's WhatsApp Cloud API: customers browse services, see required documents, fees and processing time, and request a staff callback.
• Designed the staff follow-up workflow: one open enquiry per phone number, New, Today and Overdue tabs, tap-to-call and WhatsApp buttons, and call outcomes that set the next follow-up date.
• Secured the webhook by verifying Meta's request signature on the raw body, de-duplicated retried deliveries by message id, and kept the conversation logic a pure function with no I/O.
• Kept running cost near zero by replying only inside customer-opened conversations with no paid message templates, and read the service catalogue live from Supabase so staff edits reach the bot.
• Built a CV-tailoring app that turns a hand-verified fact bank and a job description into a one-page tailored CV PDF, with the LLM limited to selecting and lightly rewording bullets.
• Enforced fact checks in code: any number, skill or bullet id not backed by the fact bank is rejected, followed by a spacing lint, one-page check and rank-based overflow fitter before export.
• Added on-demand fit analysis and cover letters: JD requirements marked met, partial or missing with fact-bank evidence, a match score computed in code rather than by the model, and token cost per run.
• Built the application tracker on Supabase: stored PDF and JD per application, automatic status history, resumes locked once sent, a follow-up queue and response-rate insights by CV angle and source.

Outcome