Systems over heroics
If it only works while I'm in the room, it isn't built yet. The test of a system is what happens on the day its author is unavailable.
Loading local time
Not finished. Just beginning to build what matters.
I'm Suraj. I work with numbers for a living and with systems by instinct — statistics, applied AI, agriculture, and the unglamorous machinery that turns an idea into something that keeps running after you leave the room.
This is not a portfolio of finished things. It's a working site: what I'm learning, what I'm building, and what I've got wrong so far, kept in public on purpose.
01 How I work
Six rules I actually use
Not aspirations. These are the filters an idea has to survive before I give it my evenings.
Measure what you'd defend
Working in official statistics teaches you fast: a number is a claim, and someone will eventually make a decision on it. Collect what you'd be willing to explain line by line.
Compound beats spike
A thing that improves one percent a week outruns a launch that peaks and decays. Most of my choices are just an argument for staying in the game longer.
Stay close to the ground
The person doing the work knows things the dashboard doesn't. Farmers, operators, field staff — the design gets better every time I ask before I assume.
Publish before it's ready
Half-formed work invites correction; finished work invites applause. Correction is worth more. This entire site is that rule applied to itself.
Health is infrastructure
Sleep, strength and attention are not rewards for finishing the work. They're the equipment the work runs on, and they depreciate quietly.
02 Domains
Six fields, one question
How do you make a good thing keep happening without you? Each of these is a different attempt at the same answer.
fig. 01 — neural corridor · 680 nodes in view · no end
Building with
Artificial intelligence
The interesting part of AI right now isn't the frontier — it's the floor. One person with a model and a clear problem can do work that used to need a small department. That changes who gets to build things.
My interest is applied and narrow: agents that clear the boring backlog, models that read the documents nobody has time to read, and interfaces where a non-technical person can ask a real question of their own data.
- Retrieval over trust — cite the source or don't answer
- Small, evaluated tools beat one large unpredictable one
- A human stays on the hook for every consequential output
fig. 02 — a distribution, as terrain · running to the horizon
Working in
Statistics & measurement
My day job is official statistics. It's slower and more consequential than most people assume: the design of a questionnaire decides what a country is later able to see about itself.
The gap I keep circling is between a number and a decision. Most published data never reaches the person who could act on it, in a form they could act on. That distance is a design problem, not a data problem.
- Sampling and non-response — where the truth actually leaks
- Making public data legible to the people it describes
- Small-area estimation for places too small to be seen
fig. 03 — one field, from inside it · one season
Learning
Agriculture
Farming is the oldest system still being actively rewritten, and the one where a bad model costs somebody their year. What draws me is the economics of the small holding — where an extra input has to earn its cost within one season.
Cheap sensing and better forecasting are finally arriving at that scale. The open question is whether the advice reaching a farmer is any good, and whether it's specific to their soil or to an average that describes nobody.
- Low-input systems that survive a bad year
- Soil as a balance sheet, not a consumable
- Water accounting at the plot level
fig. 04 — load and recovery · the axis drifts upward
Practising
Health & training
Strength training is the clearest model of compounding I've found in the physical world. You apply a stress, you recover, you adapt — and the only way to lose the gains is to stop showing up.
I treat sleep and training as fixed costs rather than optional ones. Almost every bad decision I can remember was made on a bad night's sleep.
- Progressive overload, tracked honestly
- Recovery is where the adaptation actually happens
- Consistency at seventy percent beats perfection you can't sustain
fig. 05 — attention · 460 particles gathering and letting go
Practising
Attention & inner work
I keep a contemplative practice, and I've stopped treating it as separate from the technical work. Both are training the same faculty: staying with one thing long enough for it to reveal what it actually is.
The practical payoff is unglamorous. Fewer reactive decisions. More tolerance for a problem sitting unsolved for a week.
- Attention as the scarcest input I have
- Silence as a debugging tool
- Detachment from the outcome, not from the work
fig. 06 — one percent weekly, compounded · drawn in space
Building toward
Growth & innovation
Innovation gets described as a flash. In everything I've watched closely it looks more like a boring curve that stays flat for an embarrassingly long time and then stops being flat.
So I optimise for the part most people skip: staying interested through the flat section, and keeping experiments cheap enough that I can afford to run a lot of them.
- Many small bets, capped downside
- Write the prediction down before the result
- Kill it fast, or commit for years — nothing in between
03 What I'm building
Sown, not yet standing
Honest stages. Nothing here is claiming to be a company. If one of these is your problem too, that's a good reason to write.
Prototype
Ask-your-data
A plain-language layer over public datasets, so a question like "how did rainfall move in my district over ten years" returns a chart and its source, not a hunt through portals.
Needs a clean pipeline for one state's data as a proof.
Sketch
Plot ledger
A season-long record for a single farm — inputs, water, labour, yield — simple enough to survive contact with an actual field, and useful enough that next season's decisions get sharper.
Needs two or three farmers willing to break version one.
Idea
Learning that sticks
Teaching statistics the way it's actually used — starting from a decision someone has to make and working backwards to the method, rather than starting from formulae and hoping for relevance.
Needs one course written end to end before it means anything.
Running
These notes
This site, updated as things change. It's the cheapest experiment I run and the one that has already returned the most: it makes me finish thoughts.
Needs nothing but the habit.
Idea
A fund of time
A deliberate share of each year set aside for work with no expected return inside five years. Treated as a real allocation, and protected like one.
Needs discipline more than resources.
04 Notes
Thinking out loud
Short pieces, written to find out what I think rather than to announce it.
The number is the easy part
Every statistic arrives with an invisible attachment: the assumption that made it collectable. Change who you asked, when you asked, or what counted as an answer, and the number moves without anything in the world moving. I've learned to read the method first and the headline second.
Leverage arrived before we were ready
A model can now draft, summarise, translate and search on demand. The constraint has shifted from capability to judgement — knowing which problem deserves the effort at all. That's an old skill, and no tool ships with it.
The flat part is the whole job
Compounding curves are misleading in retrospect. In the moment they feel like nothing is happening, for years. The only real skill is arranging your life so that staying is affordable.
05 Write to me
The day ends with a conversation
I read everything and I answer most of it. Ideas, collaboration, a sharp disagreement, a dataset you think I'd like — all of it is welcome.
Email seekingsuraj@gmail.com- You're working on something in statistics, agriculture or applied AI
- You think one of the rules above is wrong
- You want to break one of the prototypes