Case
Studies

Problem

A medical transcription operation received audio from several clients over several FTP servers and managed everything by hand across Drive folders. Handoffs between intake, QA and delivery were slow and error-prone.

What I built

• •Scheduled FTP ingestion from multiple clients and multiple servers into the right Google Drive folders, with no manual copying.
• •On arrival of each audio file, a background automation creates a job folder with the transcription and the correct doctor template, chosen from the file name and doctor folder. New doctor templates are added without code changes.
• •Role-based dashboards: admin, manager (assigns files to QA and approves), QA reviewer (checks transcripts), and employee.
• •Approved batches go back out through Drive to the client FTP, including the case where files arrive from one FTP and must be delivered to a different one.

Outcome

Intake, assignment, QA review, approval and delivery run as one pipeline across multiple clients and servers, with each role seeing only its own work.

Problem

A government cyber-complaint team received a high volume of complaints that were sorted by hand, which was slow and inconsistent.

What I built

• •n8n workflow that reads each complaint and classifies it with GPT-4.1-mini into fixed categories.
• •Loop plus Wait throttling so large batches stay inside LLM rate limits instead of failing midway.
• •Structured output so every result lands in the same schema for downstream routing.

Outcome

Manual first-pass sorting replaced by an automated flow that handles whole batches without babysitting.

Problem

Day-to-day operations relied on repetitive manual steps across several internal processes.

What I built

• •Worked on an internal SaaS product that automated much of that manual work.
• •Built supporting internal tools and automations around it.

Outcome

Repetitive manual work moved into software. Details are internal to the employer, so they are summarised here.

Problem

Important mail was getting buried across four separate Gmail accounts.

What I built

• •Scheduled n8n workflow pulls the last day of mail from all four accounts and merges it.
• •An LLM step sorts it into Needs action, Worth a look, FYI and Skipped. Email text is treated as untrusted input.
• •One HTML digest is sent by email each morning.

Outcome

One email each morning instead of checking four inboxes.

Problem

Calls missed while busy meant lost context and no record of what the caller wanted.

What I built

• •Unanswered or busy calls forward to a voice agent on Plivo that talks in Hindi and Telugu.
• •n8n webhook fires notifications. Recordings and transcripts are stored in Supabase.
• •Dashboard with Supabase Auth, installable as an app, with LLM call summaries.

Outcome

Every missed call is answered, recorded, summarised and pushed to me.

Problem

Small businesses run finance and support over WhatsApp, with no automation behind it.

What I built

• •Two agents on WhatsApp Cloud API, n8n and an LLM: one for invoicing and finance, one for support.
• •Built invoice, receipt, reminder and quotation templates.

Outcome

I stopped it on purpose. Accuracy risk, per-message pricing changes and a one-number-per-client setup made the economics weak. Takeaway: validate unit economics and platform dependence before scaling a build.