Workflow mapping
The manual process written down step by step, with what each error would cost.
Service · AI & workflow automation
AI workflow automation that runs unattended.
I automate manual, document-heavy workflows with LLMs: reading emails and documents, extracting what matters, and writing it into your systems. Built to run without someone watching it, and to fail safely when the model is wrong.
- n8n
- Anthropic Claude
- LangChain
- Node.js
- TypeScript
- PostgreSQL
01 — The problem
Demos are easy. Production isn't.
An LLM prototype that handles ten sample emails is an afternoon's work. Running it on every email, every day, is different: odd formats, missing fields, a model that is confidently wrong, and a token bill nobody budgeted for.
My rule is that the model interprets and deterministic code decides. Inputs are validated, outputs are checked against your data before anything is written, low-confidence cases go to a person, and costs are capped. It's the same split Theliv is built on: detectors measure, the model only explains.
02 — What you get
What you get.
Prompt & tool design
Prompts, structured outputs and tool calls designed and tested on your real data.
Orchestration
The workflow built in n8n or as a Node.js service, connected to your email, documents and APIs.
Guardrails & fallbacks
Validation, confidence checks and a human-review queue for the cases the model shouldn't decide.
Cost controls
Model choice, caching and limits, so the token bill stays predictable.
LLM features in your product
AI explanations or summaries inside your app, clearly labelled and never presented as fact.
03 — Process
How a project runs.
One point of contact, from planning to a deployed product your team owns.
01
Map the process
Sit with the people doing it today, and collect real examples, including the ugly ones.
02
Deterministic first
Automate everything rules can handle; use the model only where judgment is needed.
03
Shadow run
Run alongside the manual process and compare results before anything goes live.
04
Go live with limits
Launch with review queues, alerts and cost caps, then loosen them as accuracy is proven.
04 — Proof
Work that shows it.
Products I've designed, built and run myself. Each has a full case study.
Theliv — AWS cost-monitoring SaaS
Claude writes the likely cause and fix for each AWS cost spike, labelled as AI beside the measured numbers. If the model is down, the measured alert still ships.
Read the case study05 — Who it's for
A good fit if you are…
Operations teams re-typing data from emails, PDFs or forms into another system.
Products that want LLM features without letting the model make unchecked decisions.
Teams whose AI prototype works in a demo but not on real volume.
06 — FAQ
Before you get in touch.
Which models do you use?
Whichever fits the task and the budget. I've shipped with Anthropic's Claude, and I keep the workflow independent of one provider so the model can be swapped.
What happens when the model gets it wrong?
Outputs are validated before anything is written. Anything that fails a check, or that the model is unsure about, goes to a person instead of into your system.
Why n8n?
It makes the workflow visible and editable by your team, and runs self-hosted. When a workflow outgrows it, the same logic moves into a Node.js service.
08 — Contact
Let's Build Something.
I take on a small number of projects at a time, from full product builds to deployment, performance and automation work: scoped, shipped, and handed over clean.
Bengaluru, India → Open to RelocationReplies under 24h