Engagements

AI work for places where being wrong is expensive.

I'm Achal Dixit. I've spent six years shipping AI inside pharmaceuticals, clinical development, logistics, and genomics: environments where a model has to be right, explainable, and able to survive an audit. I build and ship AI products, agentic systems included, in a market where ten demos die every week. Getting one of them to live in production is the whole job, and it's the part almost everyone underestimates.

What I take on → See the work
01 Fit

Worth a conversation

If none of these describe your situation, we probably don't need to talk. If two or three of them do, we should.

02 Engagements

Three shapes of work

Each is scoped in writing before it starts: fixed shape, fixed timeline, defined deliverables. Scope and terms are agreed directly; just tell me the problem first.

011–2 weeks

AI Sprint

One decision, answered properly.

A short, intense look at a single question. Is this feasible with the data you actually have? Build or buy? Why has the pilot stalled? What breaks the day this meets real users? I read the code, the data, and the constraints, then tell you what I actually think.

You get
  • A written recommendation with the reasoning laid out, not just the conclusion
  • A target architecture sketch, plus the alternatives I rejected and why
  • A risk register: what breaks, how likely, and what it costs when it does
  • A sequenced plan your team can start on Monday

Right when there's real budget behind the next decision, and you'd rather not spend it finding out you were wrong.

024–12 weeks

Commissioned strategy & architecture

The roadmap, and the architecture underneath it.

Deeper work, delivered as a defined commission. Where the capability honestly sits today, what to build and what to buy, how it gets evaluated, how it gets governed, and in what order. Written for two audiences at once: the engineers who will build it and the committee that has to approve it.

You get
  • A capability assessment grounded in your systems and your data, not a maturity questionnaire
  • Target architecture with explicit build/buy calls and the reasoning behind each
  • An evaluation strategy: what “working” means, measured, before anything ships
  • A governance and controls model: data boundaries, human review, audit trail
  • A phased roadmap with owners, dependencies, and timelines that aren't fiction

Right when you're committing serious budget and it needs to survive an engineering review and a steering committee on the same slide deck.

03Engagement-scale

Enterprise deployment

From working in a notebook to working under audit.

Hands-on build. I architect the system, write the hard parts myself, and work alongside your engineers so it is genuinely yours at the end. Retrieval and orchestration, the evaluation harness, guardrails, data boundaries, deployment, and the human process wrapped around all of it.

You get
  • A production system that is deployed, monitored, and handed over, not a pilot
  • An evaluation harness your team can run on every change after I'm gone
  • Guardrails and audit trails designed in from the start, not retrofitted under pressure
  • Documentation written to hold up under validation
  • Knowledge transfer, so the system outlives the engagement

Right when the prototype already proved the idea, and the genuinely hard part is still entirely ahead of you: making it safe, evaluated, and operable.

03 How I work

Four things you can hold me to

Evaluation before enthusiasm
The harness gets built first. A demo proves a thing can happen once; an evaluation tells you how often it doesn't, which is the number your risk function will eventually ask for.
Designed for review
Data boundaries, human-in-the-loop where it actually matters, and explanations that hold up in front of an auditor. Retrofitting compliance costs far more than building for it.
Production, or don't start
I've put ten-plus AI systems into environments with real consequences. Almost all of the interesting engineering happens after the demo, and that's the part most engagements skip.
A straight answer, including no
If the honest recommendation is “don't build this” or “buy it instead,” that's what you'll get, in week one. It's the cheapest deliverable I can hand you.
04 Evidence

Measured, not asserted

Figures are as reported by the organisations involved. Where the work is sensitive or ongoing, the client stays at domain level.

$4.5M
External vendor spend removed
GenAI for clinical study reports · Bristol Myers Squibb
$10M
Vendor cost avoidance
Risk-based monitoring in R&D · global pharmaceutical company
~50%
Audit-preparation time cut
Inspection & regulatory-readiness intelligence · BMS · Science & Innovation Award 2025
46%
Faster first regulatory draft
Compliance-native retrieval · 12-week validated pilot, clean on data boundaries across ~9,500 operations

The research and the code are both public: two papers in submission on trustworthy AI evaluation, and vibebill, an open-source cost tool shipped on npm. Read them before you call. Full archive →

05 How it starts

Three steps, no funnel

01
Tell me the problem
An email, a few paragraphs. What's stuck, what you've already tried, and what “done” would look like. Enough detail that I can think about it before we speak.
02
Thirty minutes
A call where I tell you whether I'm the right person for this. Sometimes I'm not, and when that happens I'll say so and point you at who is.
03
A one-page scope
Shape, timeline, deliverables, and what falls outside it, all in writing, before either of us commits. No proposals with an appendix.

Tell me what's
actually stuck.

Selective availability. I hold a very small number of engagements at a time, alongside my full-time work. That's a real constraint, and it's also the point: I only take on work I can do properly.

If the problem is hard, the stakes are real, and being wrong is expensive, that's exactly the kind of thing I want to hear about.

Emailachaldixit22@gmail.com
Typical replyWithin two working days
Start a conversation LinkedIn ↗ See the work