Software for established businesses

Sort
it out.

Everyone is telling you to put AI in your business. Nobody mentions that it can’t read the information you already have.

Tell me what’s wasting your week Book a 20-minute call
  • 2 days → 10 minutes to pull a report
  • 1–2 months to build, each
  • 0 have been switched off

Work

Built, delivered, still running.

Five systems, all still running — and the first two are this page’s argument, in order. Get twenty years of records out of an Access file and into a shape software can read; then AI over a company’s own data stops being a demo and starts answering. I built each as the engineer on the inside rather than a supplier delivering to a brief, so every decision was mine and I can tell you why. The businesses are described rather than named: these hold client records under confidentiality.

  1. Billing & records

    IP law practice, New Delhi · 2026 · built end to end

    In daily use
    • Data migration
    • Schema design
    • ERP deployment

    145,000 records moved. Not one lost.

    The whole firm’s practice data moved into something they open in a browser, and for the first time into a shape software can read. Twenty years of it, normalised once and loaded into an ERP they now run on. This is the boring half, and nothing intelligent works without it.

    What it was like before

    Twenty years of a practice lived in one Access file that a single person could have open at a time. Deadlines were kept in a parallel spreadsheet, because the database couldn’t be trusted to remind anyone. Every clearance report was assembled by hand from both.

    What I did

    Rebuilt the schema as 28 normalised tables, then moved 145,000 records into it with a migration that ran the old and new systems side by side until the counts matched on every table. On top of that I built a replacement billing application. What the firm needed turned out to be wider than billing. Cost-centre allocation and firm-wide reporting are ERP work, so rather than grow a billing tool into a homemade ERP, we moved them onto an open-source one. The schema work carried over unchanged. The records went straight in.

    The hard part

    I recommended an ERP in the first week. They said it was too big and too expensive, so I built the billing system they’d asked for — and within a month their own feature requests had outgrown it. What changed their mind was watching that happen, not me arguing a second time; and the ERP we landed on was open-source and self-hosted, which answered the objection they’d actually had.

    records moved
    145,000
    tables rebuilt
    28
    days of downtime
    0

    Built end to end · schema and migration in 1 month · running since August 2026

    SOURCE SYSTEM
    Schematic. Screens withheld — client confidentiality.

  2. Company knowledge base

    AI company, Gurugram · 2026 · consultant

    In daily use
    • Search over your own data
    • Meeting follow-up
    • Reporting

    Everything the business already knew, in one place that could answer.

    Email, documents and meeting notes pulled together into a system that answers questions about them — and says so plainly when it can’t. The documents and the index stay theirs; which model reads them is a setting.

    What it was like before

    The company knew a great deal and could not retrieve any of it. What a client had said in a meeting eight months ago sat in someone’s notes; the decision that followed sat in email; the document it produced sat in a drive. Answering one question meant asking three people who were in the room.

    What I did

    Pulled email, documents and meeting transcripts into one index, and put retrieval over the top of it so a plain question gets an answer with its sources attached. The part that took the longest was teaching it to say I don’t know — a system that invents an answer about your own business is worse than no system.

    The hard part

    The first version was confidently wrong in two different ways: it invented relationships between people and companies, and it got plain, checkable facts wrong while sounding completely certain. Relations we fixed with a classifier over subject-verb structure, which never came out elegantly and still hasn’t, but is right often enough to be trusted. For the plain facts we stopped asking it to remember at all, and grounded them against an external model instead.

    meetings → 92 written insights
    108
    tasks people went and did
    304
    users across 121 organisations
    64

    Built as consultant · 1 month · running since February 2026

    SOURCES ANSWER
    Schematic. Screens withheld — client confidentiality.

  3. Mission planning aid

    Drone operator · 2025 · developer

    Passed acceptance
    • Geospatial scoring
    • Route planning
    • Decision support

    A planning score that says where it stops being trustworthy.

    A planning tool that ranks options across terrain, and ships with a written account of what it must never be the only input to.

    What it was like before

    Routes and sites were assessed by hand, on maps and in spreadsheets, one option at a time, with no consistent way to compare them.

    What I did

    A scoring model over terrain, structures and line of sight that ranks the options instead of listing them. Built over two months of feedback rounds with the people who would actually use it, then put through the client’s own acceptance testing at 90% accuracy.

    The hard part

    The first version of the scoring model was confidently wrong in six ways at once, and none of them looked like errors: thresholds that were absolute instead of scaling with the search area, cliff effects where a metre either side of a boundary changed the answer, and a terrain input the system fetched on every run and then never used. A score has no way of looking wrong, so I wrote an operational review against my own tool — what it was adequate for, and what it must never be the sole basis for — and handed that over with it.

    accuracy at acceptance
    90%
    scoring faults found and fixed
    6

    Delivered 2025 · 2 months · passed client acceptance testing

    TERRAIN RANKED
    Schematic. Screens withheld — client confidentiality.

  4. Operations control room

    Drone operator · 2025 · developer

    Delivered Dec 2025
    • Live tracking
    • Video feeds
    • Desktop app

    Live positions and video on one screen, under five seconds behind.

    A control room display for flights in the air: where everything is, and what it is seeing, in one place.

    What it was like before

    Positions were known by radio. Where everything was at any moment was held in people’s heads rather than on a screen.

    What I did

    A map-based control room carrying live positions and video at under five seconds of delay, shipped as a desktop application for the operators who run it.

    The hard part

    The software was the easy half. There was no video to show until the cameras were physically out where they were needed and installed, and getting hardware distributed runs on a much slower clock than code. Most of the schedule was that, not the screen.

    delay on live video
    <5s

    Delivered December 2025 · 1 month

    FIELD CONTROL
    Schematic. Screens withheld — client confidentiality.

  5. Stock tracking

    Drone operator · 2025 · developer

    Delivered 2025
    • Inventory
    • Workflow automation

    Inventory that knows what is on the shelf without anyone walking to it.

    The unglamorous one: stock for fifty-plus pilots and engineers, counted by software instead of by a walk to the store.

    What it was like before

    Stock for fifty-odd pilots and engineers was counted by walking to the store and looking.

    What I did

    An inventory system that tracks what is issued, returned and on the shelf, so a count is a query rather than a trip.

    The hard part

    The problem was the input, not the software. What needed tracking wasn’t being recorded anywhere; there was no field for it at all. Before anything could be counted, the store had to start writing it down.

    less time counting stock
    40%
    pilots and engineers
    50+

    Delivered 2025 · 1 month

    STORE COUNT
    Schematic. Screens withheld — client confidentiality.

How it works

Four steps.
No proposal document.

You’ll know what it costs before anything is built, and you’ll own everything at the end of it.

  1. 01

    We talk

    You describe what’s annoying. I ask awkward questions about where your information actually lives. Twenty minutes, no charge, and no obligation at the end of it.

    Free · 20 minutes · no proposal document

  2. 02

    One piece, priced first

    We pick a single job and I quote a fixed price to build just that, on your real information. If it doesn’t work, you’ve spent a fortnight finding out instead of a year.

    Fixed price · agreed before anything is built · you own what it runs on

  3. 03

    Put to work

    It goes live where your team already works, with the dull essentials handled — logins, backups, a record of what happened, and someone to call when it misbehaves.

    Live on your systems · nothing switched off early

  4. 04

    Yours

    The software, the accounts and the passwords are in your name and written down. Keep me on monthly if it’s useful — never because leaving would be painful.

    Your accounts · your code · documented

And then it’s boring

Sorted.

Start with a conversation

Questions

The ones I’m always asked.

What does it cost?

A fixed price per piece of work, quoted before anything is built, so there is no meter running. Ongoing support is a monthly figure you can stop at any time. The first conversation is free and doesn’t end in a proposal document.

Price for the first piece
Fixed, agreed first
Ongoing support
Monthly, stop anytime
Accounts, passwords, code
Yours
First conversation
Free

The figure depends on the job, so I’d rather quote yours than publish a range you’d have to argue yourself out of. If it’s obviously more than you want to spend I’ll tell you then, not three weeks later.

Yours isn’t in here? Ask me something else

Things I will
happily not do

Rebuild a website that works perfectly well because you saw a nicer one.

Add artificial intelligence to something that isn’t broken, so that it becomes broken and expensive.

Charge you for a research phase whose only deliverable is a document.

Put your business’s information in an account you can’t log in to.

Call something finished while it only runs on my computer.

Get in touch

Tell me what’s
wasting your week.

Bring the messy version. Almost every project here started with somebody saying “there must be a better way to do this” — you don’t need to know what the answer is.

manastanejaa@gmail.com Book a 20-minute call

Based in New Delhi. I work with businesses across Delhi NCR in person, and elsewhere over a call. Replies within one working day.