One task your team repeats every day, automated with the simplest tool that works.

I work out which repetitive task is worth automating, then build it with the simplest tool that does the job. Sometimes that is a script. Sometimes it is a hosted language model. When your data must stay on your own machines, it is a small model fine-tuned on your examples, under one billion parameters, small enough to run on a laptop or a modest server. I connect the result to the tool where the task happens. It starts with a consultation, and I only build if automating the task is worth it.

Where it fits

The tool sits between the work your team already does and the software it already uses. It handles one task and passes the result on. Which tool it is depends on the task, and I choose the simplest one that does the job.

How the tool fits Three connected steps: your task, done by hand today, goes to the simplest tool that works, which connects to one tool you already use. Your task Done by hand today The simplest tool that works A script, a hosted model, or a small model trained on your examples Tested on examples it has not seen Your workflow Connected to one tool you already use How the tool fits Three connected steps: your task, done by hand today, goes to the simplest tool that works, which connects to one tool you already use. Your task Done by hand today The simplest tool that works A script, a hosted model, or a small model trained on your examples Tested on examples it has not seen Your workflow Connected to one tool you already use

The right tool for the job

A script or a set of rules

If the task follows fixed steps, plain automation is usually cheaper, faster and easier to check than a model. I will say so rather than sell you one.

A hosted language model

If the task is messy or needs broad understanding, and the data is fine to send to a provider, a hosted model with careful instructions and a test set is often enough. You pay per request and need no special hardware.

A small model of your own

If the data must stay on your own machines, the volume is high, or per-request fees add up, I fine-tune a model under one billion parameters on your examples. It runs on a laptop or a modest server, and I test it on examples it has not seen before handover.

A mix

A hard task is often several small steps. A script or a small model handles the easy ones, a hosted model handles the hard ones, and a person approves the result.

See the task

Example email in

Hi, my invoice from last month shows the wrong amount. Can you correct it?

What the tool returns

{"category": "billing", "needs_human": true}

Illustrative example, not a client result.

Tasks worth automating

Sorting incoming emails or support requests into the right category.

Pulling named fields, such as dates, amounts and names, out of invoices, forms and messages.

Drafting standard replies that a person reviews before sending.

Tagging and routing records in a spreadsheet or database.

Copying data between two systems that do not talk to each other.

Checking records against a list of rules and flagging the ones that fail.

Is it a fit

A good fit

The task repeats every week and a person does it by hand.

You have past examples of it done correctly, or the steps can be written down as clear rules.

A wrong answer can be caught by a person or a simple check.

Probably not a fit

Open-ended questions with no single right answer to check against.

Long documents that need careful reasoning across many pages.

Tasks that nobody can describe in rules or show examples of.

Three ways to work together

Workflow review
Two to three working days, fixed price. I go through the tasks your team repeats, say which are worth automating, which tool each one needs, and what building it would involve. You get a short written report.
Two-week pilot
One task automated, tested and connected to a tool you already use. The details are below.
Monthly support
After handover, I check the results, fix errors and update the model or rules as your data changes. Agreed month by month, and only if you want it.

A two-week pilot

Scope
One task, one tool, one integration.
Timeline
Two weeks from the day the scope is agreed.
Price
Fixed, quoted in writing after the first call.
Not included
General-purpose chatbots. Tasks that need open-ended reasoning or long documents. Hosting and maintenance after handover unless agreed separately. Any accuracy promise beyond what the test report shows.

How it works

  1. Quote: You send a request. I reply within two working days with questions or a fixed quote.
  2. Day 0: A 30-minute call to confirm one repetitive task worth automating. If a small model is the wrong tool, I will say so and suggest a simpler one, or tell you plainly that the task is not worth automating.
  3. Days 1 to 3: You and I agree how success is measured, and I prepare your examples or write down the rules the task follows.
  4. Days 4 to 10: I build the tool and test it on examples it has not seen.
  5. Days 11 to 14: I hand over the tool, a plain-language accuracy report, and instructions to run it.

Questions

Do I need to be technical?
No. I need examples of the task and a description of where it happens. I handle the build and connect the result to one tool you already use.
Where does my data go?
It depends on the tool. A script or a small model runs on your own machine. A hosted model sends the text of each request to its provider, so I only suggest one if that is acceptable to you. I will agree in writing how your examples are handled, stored and deleted before you send anything.
How accurate will it be?
I will not promise a number in advance. After building it, I test the tool on examples it has not seen and give you a plain report of what it gets right and wrong, so you can decide whether to use it.
Do I need to buy special hardware?
No. A small model runs on a laptop or a modest server. I do the training on my own machine, so you do not buy a GPU.
What if a small model is the wrong tool?
Then I will tell you on the first call and suggest a simpler or different tool, or say the task is not worth automating. You will have spent 30 minutes, not two weeks.
What happens after the two weeks?
You receive the tool, the report and instructions to run it. Hosting and monthly support are separate, and only if you want them.

About

I'm Prosun Banerjee, founder and CEO of Lemnar. I build machine learning systems in Python and PyTorch, and I automate the workflows around them. I choose the simplest tool that does the job, even when that is a plain script. I like small models because you can test them properly, read exactly what they get wrong, and run them on hardware you already own. I work on my own, so you speak directly to the person who builds and tests your tool.

Tell me which task your team repeats most.

Request a quote