Independent engineer · Remote, UTC+3 · Taking work now
Your team is losing hours to work a machine could do. Here's roughly how many.
Move the sliders. That number is yours, calculated conservatively, and I've listed everything it deliberately doesn't count. If it comes out small, don't hire me — I'd rather you know that in ninety seconds than four months.
↓ Open the calculator at the bottom of your screen
Where it actually goes
Four leaks, and how you know you've got one
Nobody has a line item called "looking things up". It hides inside four ordinary-looking activities.
Finding what's already written down
The answer exists — in a policy doc, an old ticket, a contract, a wiki page nobody has opened since 2023. The cost isn't knowing it; it's locating it, and then trusting that you found the current version.
The tell: people ask a colleague instead of searching, because asking is faster.
Deciding what each new thing is
Tickets, emails, applications, invoices, incident reports — read, categorised, prioritised, routed. Individually trivial, collectively an hour a day of someone senior enough to judge.
The tell: the first hour of the morning is spent working out what the day is.
Retyping what a document already says
Numbers and fields moved by hand from a PDF, form or contract into a system that needed them structured. Slow, and the error rate stays invisible until it's expensive.
The tell: two windows side by side, one being read, one being filled.
Repeating a process that's 80% the same
Several steps with real judgement somewhere in the middle: draft it, check it against something, look up the account, escalate if the numbers disagree. Forty minutes, most weeks, mostly identical.
The tell: there's a doc explaining how to do it, and it's out of date.
Before you trust that number
What the calculator deliberately doesn't count
Every vendor ROI calculator is built to flatter. This one is built to survive your CFO, which means being honest about its own weaknesses.
- It assumes 46 working weeks, not 52 — leave, holidays and the weeks where nobody does anything useful.
- The automatable share is a guess until it's measured. That slider is the single biggest source of error on this page. Nobody, including me, knows the real figure for your process until a pilot has produced one.
- Recovered hours are not automatically recovered money. Unless you reduce headcount or the freed time goes to work that earns, the saving is capacity, not cash. Capacity is usually worth more — but it's a different conversation with your finance team.
- It ignores ramp-up. People are slower with a new tool for a few weeks, and some never adopt it. Budget for that.
- It excludes model and infrastructure costs, which you pay directly to your provider. For most processes this is tens to low hundreds a month, but it isn't zero.
- It says nothing about error rates. A system that's fast and wrong is worse than the status quo, which is why measurement comes before deployment, not after.
How the number gets captured
Four steps, and the first one isn't code
Agree what "correct" means
A hundred real cases with the right answer written next to them, produced with the people who currently do the work. This exists before any system does, and it's what turns the automatable-share slider from a guess into a measurement.
Prepare the content
Splitting, structuring and cleaning what the system will read. Unglamorous, skipped by most pilots, and the most common single reason answers come back plausible and wrong.
Build it so it can run
Retrieval, permissions that match your existing document access, cost ceilings, latency budgets, logging of what was retrieved and why, and a deployment pipeline. The difference between a demo and a system.
Hand it over and leave
Runbook, evaluation harness, and a working session with your engineers. Everything in your infrastructure and your repositories. If it stops working when I disappear, I built it wrong.
What it costs to find out
Published, because you shouldn't need a call to learn a price
Assessment week
$3,500One week. Which process first, whether your data supports it, a written report with risks and costs. Credited in full against a pilot.
Pilot
from $7,500Two to four weeks. One process end to end on your real data, plus the evaluation set. You get a measured number, not a demo.
Fixed-scope build
$15k – $45kWritten spec and acceptance criteria. Permissions, guardrails, monitoring, pipeline, runbook, handover. You know the number before we start.
Fractional
from $5,000/moOr $800 a day. In your Slack and your tracker, two clients at a time, 30-day rolling after the first month.
A full-time senior AI engineer runs well past $150,000 a year fully loaded, plus months of hiring risk and a decent chance the role outlives the work.
When the number lies
Four situations where I'd tell you not to do this
- Nothing is written down. If the knowledge lives entirely in people's heads, retrieval has nothing to retrieve. That's a documentation problem in an AI costume, and it isn't one I can solve for you.
- Nobody can say what a correct answer is. If it's a fresh judgement call every time, there's nothing to measure against and no way to know whether the system is helping.
- Nobody will own it on Monday. Documents change, models get deprecated, questions drift. Without an owner it decays quietly and you've bought a liability.
- The recoverable figure is under about $30,000 a year. A custom build won't pay back inside a sensible horizon. Buy an off-the-shelf tool, or leave it alone — both are better answers than hiring me.
Who's doing the work
I came from the side where things break at 3 a.m.
An automated trading platform placing orders on the Texas wholesale power market. A US equities trading platform. A national payment gateway I helped break out of its monolith. Kubernetes clusters and data pipelines under all of it.
None of that was really about writing code. It was about knowing what to do when something drifts — metrics on every stage, alerts tuned to catch an anomalous run rather than a merely slow one, and enough logging to explain afterwards what happened. An LLM inside a business process is that same problem wearing new clothes.
Python, Go, Kubernetes, Docker, AWS, GCP, Postgres, MongoDB, Redis, RabbitMQ, Celery, Prometheus, Grafana — and 76 technical posts on DEV.
Asked every time
The four questions I always get
Isn't this just n8n or Zapier?
If n8n solves your problem, use n8n — it's cheaper than me and running this afternoon. Those tools move structured data between apps on rules you write in advance, and they're good at it.
My work starts where the rules run out: unstructured content, judgement in the middle, and answers that have to be checked rather than assumed. That needs retrieval over your own material, a way to measure whether it's right, permissions so it can't surface what someone shouldn't see, and monitoring for drift. None of that is a node on a canvas.
What happens to our data?
Decided in week one and written into the agreement: enterprise API tiers with no-training guarantees, a specified region, self-hosted embeddings and vector storage inside your own infrastructure, or fully local models where the data cannot leave at all.
Access control matters as much as storage — retrieval has to respect the permissions your documents already carry.
How accurate does it actually get?
I won't quote you a number before seeing your content, because anyone who does is guessing. What I will commit to is that you'll have a measured figure on your own data before you decide whether to deploy — that's the entire point of the pilot.
If the measured number isn't good enough to use, you'll have spent a pilot fee to find out instead of a build.
You're one person. What if you disappear?
Everything runs in your infrastructure, in your repositories, with a runbook and an evaluation harness your engineers can run without me. No proprietary layer, no hosted black box. Handover is a phase, not an afterthought.
If that isn't reassurance enough — and for some organisations it reasonably isn't — an agency is the right call and I'll say so.
Start here
Send me your number and the process behind it
That's the whole first message. Thirty minutes on a call and I'll tell you straight whether it's worth doing — I lose nothing by saying no, and you lose a great deal by hearing yes from someone who shouldn't have said it.
info@bshadmehr.me