Bengaluru, India

I'm Prakhar.

I like working on tough problems, learning from people I look up to, and improving myself. This site is where I keep my work, my writing, and the things I'm still figuring out.

For work, I'm an enablement engineer at Draup, building Etter.ai — I help companies get AI systems working inside their real business. Before that I founded a startup in AI evaluation, and before that I was mapping floods from satellite radar.

Open to forward deployed & deployment strategist roles

01Writing

What happens to knowledge when we make it measurable

I build systems that turn human expertise into data. These essays are me taking that seriously — what it's good for, and what it quietly costs.

Long read · In draft

Legibility engines: what workforce intelligence does to tacit skill

Some knowledge can't be written down — the operator who hears a bearing failing, the account manager who knows which clause the client will actually fight over. James Scott called it metis. An entire software category now exists to render it legible.

The argument of this essay is that legibility isn't neutral capture. Specifying a skill finely enough to measure it is the same act that specifies it finely enough to automate it. I build these systems for a living, which I think obliges me to be precise about the trade rather than sentimental about it.

The moment you can score a skill, you have written most of the spec for replacing it.

Essay · In draft

The firm as a vessel

A counter-reading, drawing on Schumacher, the Toyota production system, and the German Mittelstand — firms that treat building human capability as a deliberate output rather than a cost leak. If the first essay describes a curve, this one asks who has managed to sit off it, and what they gave up to do so.

Published · Report

The attenuation trap

On measurement reliability in skills assessment. When two noisy instruments are correlated, the observed relationship is pulled toward zero. Spearman's correction for it is a century old and routinely ignored by people shipping skills-match scores today. A practical piece about why your model's ceiling is often your rater's floor.

02About

A civil engineer who kept following the data

The short version, and the facts at a glance.

I studied civil engineering at BITS Pilani and spent my thesis segmenting flood extent from synthetic aperture radar — the kind that sees through cloud cover, which is the only condition under which floods actually happen. That was the turn. Every job since has been some version of the same problem: take messy real-world data, and make it answer a question someone urgently needs answered.

During India's second COVID wave I built Bedstats.in, running verified hospital resource data across a volunteer network of two hundred people. Later I co-founded Styx, a startup building evaluation tooling for agentic AI. We did 129 customer interviews and raised around $100K, and I learned the most expensive lesson available to a founder: we had the right diagnosis on the wrong timeline. I wound it down rather than push a product at a market that wasn't hurting yet.

What I took from all of it is a particular temperament. I'm comfortable being the person in the room who has read the customer's data before the meeting, who will say the demo isn't ready, and who can then go and make it ready. Most AI deployments don't fail on the model — they fail on the messy human and organisational parts — and I like the deployment end of the business because it's where claims meet consequences.

Outside work I read a lot of philosophy of technology, keep a self-hosted agent stack running on a VPS mostly for the pleasure of maintaining it, and am slowly learning French.

03Work

Where I've done it

Five years of taking technical systems to the people who have to live with them.

PresentCurrent

Enablement Engineer

Draup Inc. — Etter.ai

Etter.ai is an AI-native workforce intelligence platform. Enablement engineering means owning the distance between what the product can do and what a specific customer needs it to do this quarter.

Turn raw enterprise exports — employee-level workforce data, org hierarchies, unlabelled role taxonomies — into something the platform can reason over.
Build the demo environments that carry deals: data-backed and specific enough that a buyer recognises their own organisation rather than a sample tenant.
Design the evaluation layer. When a model infers a skill nobody wrote down, someone has to decide whether it was right — a measurement problem before it's an ML problem.
Translate in both directions: customer language into product requirements, product constraints into something a stakeholder can act on.

Deployment · Data pipelines · Evaluation design · Pre-sales

2023 — 24Founded

Co-founder

Styx — evaluation for agentic AI

Backed by Pontaq Ventures and PIEDS, roughly $100K raised, with research conversations at Meta and Google and 129 customer interviews behind the roadmap. Wound down deliberately when it became clear we were early rather than wrong. Nearly everything I know about qualifying a deployment came out of those interviews.

2023

Geospatial intelligence

Suhora Technologies

Satellite and SAR-derived intelligence products. Imagery pipelines teach a specific discipline — the data is expensive, noisy, and indifferent to your schema.

2021Crisis build

Founder

Bedstats.in

Hospital bed and oxygen availability during India's second COVID wave. Grew to over 6,000 verified case pipelines a day across 200+ volunteers, built under conditions where a stale record was worse than no record at all.

Selected projects

Skills intelligence explorer

A single-file analytical dashboard built for an enterprise RFP, fed by an employee-level workforce workbook. Moves a reviewer from an organisation-wide skills distribution down to one team's adjacency gaps without leaving the page — shipped with a published data dictionary so the client could interrogate the method, not just the output.

Self-hosted agent stack

A personal VPS running a Docker-composed agent environment: a LiteLLM proxy in front of several model providers, a set of MCP servers, and routing rules I maintain myself. Not a demo — it's how I work daily, and it's why I can discuss routing economics and tool-calling reliability from experience.

Flood segmentation from SAR

Semantic segmentation of flood extent from Sentinel-1 radar. DeepLabV3+ and U-Net compared across held-out scenes, reaching 93.25% IoU. Undergraduate thesis, BITS Pilani.

Independent consulting

Infrastructure cost modelling, technical proposals, and delivery scoping for small clients under the Flexi Roundtables banner. A useful discipline: you write the estimate, then you live inside it.

04Contact

Say hello.

Fastest by email. I read everything and reply within a day or so.

Emailkumarprakhar601@gmail.com RésuméPDF LinkedIn/in/REPLACE GitHub@flamboyant-vinci