We help offline, built and physical businesses be more efficient

AI is a tool like anything else, a vastly more powerful tool than the ones before it, but if we do not map exactly where your processes falter or the flow breaks, no tool can help. We create a value stream map, find where the operation is losing hours or money, and wire in the correct tools, with or without AI, in a way that fixes the flow and is seamless end to end.

Systems we have built

Ref

System

What it does

Status

R-01

Khet Saathi

Supervisors on a working onion, cotton and groundnut farm in Gujarat, India needed agronomy answers in their own language, live, on the field. We deployed a farm advisory AI that answers in the language it is asked, is multi-modal, and runs on WhatsApp.

Live pilot

What goes in

Messages from farm supervisors in Gujarati, Hindi or English.

What comes out

Crop disease diagnosis, irrigation and harvest timing, and spray recommendations, answered in the language asked.

Built with

WhatsApp Business Platform, Python on Vercel, Claude Haiku and Sonnet, Upstash Redis.

Worth knowing

Every diagnosis gets a second, context free review pass. A high severity case the reviewer flags does not get treatment advice, it gets escalated to a human agronomist.

R-02

Sales call analysis

More than 150 sales call recordings sat inside a platform with no API and no export feature, so nobody could look at them in aggregate to gauge scheduler performance. We found the pattern behind the recording URLs, collected the calls, transcribed them, and scored every one against a rubric.

Completed

What goes in

Call recordings held in a platform with no API and no export.

What comes out

Each call analysed against a rubric, plus a batch pass over the transcripts for the keywords and pain points that recur.

Built with

A discovered recording URL pattern for batch download, FFmpeg validation, Deepgram for transcription, GPT-4 for analysis.

Worth knowing

The phrases it surfaced went into the paid media, the ad copy and the landing page.

R-03

Job cost analysis

Proposal data sat in the CRM with no way to see what a job actually cost against benchmark. We automatically extract the proposal line items and produce the job cost metrics from them.

Completed

What goes in

The proposal object, with its line items of material and labour numbers.

What comes out

A Google Sheet with the numbers formatted against the company metrics for job profitability.

Built with

OpenAPI, Google Sheets, Google Apps Script, Python.

Worth knowing

It eliminated the data entry time, because the sheet is generated every time a proposal closes into a sale.

R-04

Daily snapshot

Disparate systems made it hard to see the daily picture. We write the proposals, appointments, contacts, invoices and campaigns to a sheet every day, so the dashboard reads from one place instead of five.

Runs daily

What goes in

A scheduled pull of the CRM data.

What comes out

A Looker Studio dashboard.

Built with

Google Sheets, Google Looker Studio, Python, GPT-4 API.

Worth knowing

It refreshes on its own and the data is read only, so the dashboard is always current and nothing it does can touch the CRM records.

R-05

Private model workstation

Some of the work should not go to a third party model at all. We set up a local workstation that serves private models over a private network, reachable from a laptop or a phone.

Completed

What goes in

API keys for the hosted models or OpenRouter, the model selection, and the Hermes and Telegram setup.

What comes out

A private, on premises assistant for lower level daily work.

Built with

Mac Mini, Ollama, Hermes, OpenRouter, open source local models, Tailscale, and a Telegram bot.

Worth knowing

Running a non frontier model costs nothing, at the price of speed and quality. For work like home office automation that trade is fine.

R-06

Inbound call qualifying

A home services company pays for every lead twice: once for the ad, and again when nobody picks up. We built a voice agent that answers the call, qualifies the job against the questions a scheduler would ask, and books the in home estimate on the calendar.

In testing

What goes in

An inbound call from someone who clicked an ad or filled in a web form.

What comes out

Scope, timeline and contact details captured to the file, and an estimate booked on the calendar. If it is a bad time, a callback inside business hours instead.

Built with

Retell AI, GPT-4.1, an ElevenLabs voice, Cal.com for availability and booking.

Worth knowing

The agent is told what it may not say. It does not give line item pricing or price breakdowns, and when a caller pushes past that it hands off to a person rather than guessing.

R-07

OpenPTS

Timing manual work by hand is slow, and two people timing the same task rarely agree. We built an open source REST API that turns video of a task into motion codes and a standard time.

Open source

What goes in

Video of a task, or a MODAPTS sequence posted to the API.

What comes out

Standard time, units per hour, a breakdown by body region, and a repetitive strain index.

Built with

MediaPipe pose detection in the browser, a REST API on Vercel. MIT licensed.

Worth knowing

MODAPTS has been used in manufacturing and logistics for over fifty years. It is a registered trademark of the International MODAPTS Association; this is an independent implementation.

Try it

openpts-api.vercel.app

How the work runs

Map the work as it actually happens

We walk the process end to end with the on-site operators and read what is recorded on clipboards and CRMs. The idea is to find what is improperly recorded or things that are inside heads but not on paper. The first thing we build, therefore, is a Value Stream Map (VSM).

Separate the happy and not-happy paths

We check what happens when a lead is not called, when we see material on the floor we think is idle, or a truck that is waiting, and we ask for the breaks in flow, the interruptions, and also exceptions and how we treat all those. There is still nothing we would rush in to deploy, but we come up with a Gap Analysis.

Decide what should run each step

Now that we know the flow of value and flow of non-value, we create three things for each step: a rule, a judgement, or a decision a person needs to make. Rules are just if-then loops. Judgement from specific knowledge and experience goes to a model, and we measure what it produces. The decision is usually easy with those two, but we create a system for enforcing them and their exceptions so we get consistency. So now, we have a System.

Baseline it and standardize

We document this with the flexibility to adapt and scrutinize it against outcome metrics so we know what to compare it to after X number of months. Then we build with overcommunication to the entire team and deploy based on how many systems we have identified, and measure results over time.

Start with one thing that is costing you hours

Tell us what the week looks like and which part of it you would not miss. We will come back with an honest view on whether this is work for us, someone else, or nobody yet.

Book a call →

Or write to us

siddhit@esteem-automation.com

Or ask an AI

Check if this is a fit

Not ready to talk? Let your AI sort one messy week.

Three copy-paste prompts for ChatGPT or Claude: sort yesterday’s messages, check a quote before it scars, turn a hallway update into a job line. We email the guide. Free. No follow up unless you ask.

Where to send it

Who runs this

Esteem Automation Labs is run by Siddhit Sanghavi. He has spent fifteen years in offline, built and physical industries, working on sites, with trades, and creating technology and processes. As an industrial engineer he has worked in auto assembly at Ford Motor Company, diamond polishing, used car auctions at Cox Automotive through Fyusion, and residential construction as a Bath Tune-Up and Kitchen Tune-Up franchisee. More about him at siddh.it and at LinkedIn

Credentials

PMP CCA-F

Privacy Terms