Amit Kohli
Recent content on Amit Kohli
Latest posts
One Row Per What?
There's no such thing as a quick question
01/02/2026 is two different days
Scaffolding for AI
If you don't have a data dictionary, start here
What are agents and skills for an agent?
Reporting is not insight
Everyone’s seen the DIKW pyramid. Nobody knows what to do with it. Here’s a version that maps to what your organisation is actually doing - Reporting, Analysis, Insight, Action - and what it actually takes to climb from one level to the next.
The problem with averages
Imagine you have one foot in boiling water and one foot in ice. On average, you’re fine. That’s the problem with averages - not that they’re wrong mathematically, but that they’re answering a different question than the one you actually care about.
How do you know when you've outgrown your tool?
The staircase doesn’t stop at Excel. Every upgrade follows the same pattern - you hit a ceiling, you learn the new thing, and then the new thing becomes your hammer. The real problem isn’t recognising when to move on. It’s that you can’t fully evaluate the upgrade until after you’ve made it. No clean answer here.
How to know when your spreadsheet has outgrown Excel
Excel didn’t outgrow you - you outgrew it. The signs aren’t in the file, they’re in the behaviour around it. Four signals that tell you it’s time to move on, and why none of them are about file size or formula complexity.
This is why your VLOOKUP isn't working
Your VLOOKUP returns #N/A even though the value is right there. The culprit is a trailing space you can’t see. TRIM() fixes it - but the real question is where that space came from in the first place. Web forms, legacy systems, copy-paste from Word. The data is telling you something about its own provenance.
Treat the AI like an employee
You got a bad output from AI and your instinct was “the AI is broken.” But would you give a new hire a task with no brief, no context, no examples - and be surprised when they got it wrong? AI needs the same things a new employee needs - context, examples, rules, domain knowledge. Your output quality reflects the quality of your onboarding.
Your AI problems are not staff problems
Someone on your team used AI, got a bad output, and caused a problem. The instinct is to find out who. That instinct kills AI adoption. No-blame culture from aviation and healthcare has a specific lesson - punishment stops incident reports, and incident reports are how you learn. The same applies to AI. Context is everything, and you can’t build the infrastructure to capture it while…
Report or Dashboard?
Everyone defaults to the dashboard. Real-time feels like accountability. But a dashboard lets everyone avoid saying what the numbers mean. A report forces the analyst to interpret, forces a hypothesis, forces a decision. The format matters less than whether someone was forced to have an opinion.
Before you build another AI agent, write a quality charter
I have 20 agents and 31 skills. Each one felt like the right call at the time. And then one day I looked up - bloat and rot. The fix isn’t more tooling. It’s one simple document that defines what good looks like before you build anything else.
What does AI actually do to your organisation?
Prompting Foundations: how to talk to chatbots
Why Flat Classification Fails
When everything is tagged at the same level, three things happen: synonyms look like different ideas, the same word means different things in different contexts, and tools get treated as peers of the outcomes they produce. This video breaks down exactly how flat structure corrupts meaning - and why it matters a lot if AI is learning from it.
Knowledge Graphs 101: Mapping 52 Weeks of Data Concepts
Knowledge graphs turn up in almost every serious AI conversation now - RAG, semantic search, ontologies. But most explanations stay abstract. This video builds one live, using this series as the data. The point isn’t the technology - it’s what you can see once the connections are drawn that you couldn’t see before.
Every dataset is an argument you didn't write down
An M&E framework built in stable democracies quietly encodes what “functioning governance” looks like. Train AI on it and apply it to fragile states - and you’ll confidently defund programs that are working. Your dataset isn’t neutral. It’s an argument about what success means in a particular context, and this video shows what happens when that argument travels…
Why you can't answer next year's questions (disaggregated data)
The question your CFO will ask next year has no answer - not because anyone made a mistake, but because the data was collected at the wrong grain. Aggregation destroys information permanently. This video explains why the people designing your data collection need to think about questions that haven’t been asked yet, and how to do that without building a data swamp.
How a 'quick Excel' becomes a $500k data project that fails
The £500k data warehouse wasn’t the solution to the bad spreadsheets. It was the consequence of them. The real problem was the absence of data lineage and ownership from the start - and buying a warehouse before you understand that makes everything worse. This video traces exactly how one Friday spreadsheet becomes three versions of “revenue” and a very expensive non-solution.
Your Excel File is Doing Three Jobs (and failing at all of them)
Every organisation has a file that quietly became a production system without anyone acknowledging it. The problem isn’t Excel. It’s that the system has undeclared architecture - and that means it will fail in ways nobody can debug. This video names what’s actually happening, which turns out to be a prerequisite for fixing it.
Shaping the future - Intro to Data Portal & Dashboard
Shaping the future of social care data - System Overview
From Fragmentation to Transparency - Using AI to Democratise Social Care Data
What is RAG and how to get it right
Growing Pains in RAG (intro level)
Growing Pains in RAG (super-techie version)
Adventures in RAG: Are Chocolate Covered Pickles Delicious?
Data is STILL Useless
Match the feeling, not the question
I had an interesting conversation with my sister that I thought would be of interest here. Full disclosure, this post illustrates a neurodivergent thought correction, so it might read weird. But oh well. We were talking about why I feel the need to correct or provide guidance to people in my life when all they wanted was connection. She expressed concern that I would be an “actually”…
Match the feeling, not the question
I had an interesting conversation with my sister that I thought would be of interest here. Full disclosure, this post illustrates a neurodivergent thought correction, so it might read weird. But oh well. We were talking about why I feel the need to correct or provide guidance to people in my life when all they wanted was connection. She expressed concern that I would be an “actually”…
Access to Justice - Lawtech Panel
Data is Useless
SDG Direction Estimation
SDG Direction Estimation
an R-user's guide to Python for data science
R is for sure the best data science language period. That being said, Python is more popular… I suspect it’s a priors thing (i.e. there are just a ton of Python users doing all sorts of stuff, which makes it more likely that companies will choose Python as their de facto language). I will also go on the record and say that Python’s NLP game is better than R’s. So little by…
an R-user's guide to Python for data science
R is for sure the best data science language period. That being said, Python is more popular… I suspect it’s a priors thing (i.e. there are just a ton of Python users doing all sorts of stuff, which makes it more likely that companies will choose Python as their de facto language). I will also go on the record and say that Python’s NLP game is better than R’s. So little by…
10 reasons why Random Forest is the only algorithm you will need
Quick disclaimer, Random Forests, or RF, don’t solve all problems. If you want to identify cats in an image, or do text stuff like figure out what political candidate mentioned “pudding” more often, or reinforcement learning, RFs can’t help you. But for regression or classification tasks with numerical data, RFs are the best tool around. What are Random Forest models?…
10 reasons why Random Forest is the only algorithm you will need
Quick disclaimer, Random Forests, or RF, don’t solve all problems. If you want to identify cats in an image, or do text stuff like figure out what political candidate mentioned “pudding” more often, or reinforcement learning, RFs can’t help you. But for regression or classification tasks with numerical data, RFs are the best tool around. What are Random Forest models?…
How are SDG Indicators interconnected?
This post is an interactive R Markdown document. View the full post with interactive network .
How are SDG Indicators interconnected?
Intro The Sustainable Development Goals (SDGs) are a set of 17 goals that were adopted by all United Nations Member States in 2015. The SDGs are a universal call to action to end poverty, protect the planet, and ensure that all people enjoy peace and prosperity. The SDGs are Goals (the “G” in SDG). The SDGs are interrelated, and each goal has specific targets that need to be met in order to…
Knowledge Graphs but, like, for Everything
Natural Language Processing for Data Science for Good
Introduction to Network Analysis
Interactive COVID map by LSOA
Across the world, the COVID 19 pandemic has illustrated how important it is for governments to have the ability to gather, action, and disseminate new case data as close to real time as possible. Luckily in the UK, there is a reasonably good flow of information about new cases. The government has released an interactive map of cases, and they do keep it up to date. This map uses the Middle Super…
Interactive COVID map by LSOA
Across the world, the COVID 19 pandemic has illustrated how important it is for governments to have the ability to gather, action, and disseminate new case data as close to real time as possible. Luckily in the UK, there is a reasonably good flow of information about new cases. The government has released an interactive map of cases, and they do keep it up to date. This map uses the Middle Super…
Introducing CarbonFootprintR, the cool way to figure out the carbon footprint of your in-person event
This post is an interactive R Markdown document. View the full post .