Production AI is a different job from the demo.
Agents that call real tools, RAG over your own data, voice that has to answer in real time: they all work on stage and then meet customers, traffic, edge cases and a budget. I do the job after the demo, as the senior technical owner on your team or on one piece of work with a clear end.
Who this is for
- Founders with an AI product in front of customers, or about to be
- CTOs who want a senior architect beside them on the AI calls
- Investors who need to know what is really built before the wire
Ongoing
I join as a Fractional CTO, an AI Architect, a Head of AI, an AI Solutions Architecta Fractional CTO
Part-time, embedded, for as long as it helps. The title is yours to choose; the job is the same: the hard technical calls, taken with you and written down, and a team that knows where it is going.
What I own
The calls that decide whether the system holds.
- Architecture, and the model and vendor choices
- Delivery: what ships, when, and whether it clears the bar
- Quality and security: evals, guardrails, the answers buyers ask for
- The decision log your board can read
The rhythm
Stated before it starts, not discovered.
- A fixed slice of every week, in writing
- A weekly call with the team, a line to me between calls
- Written answers to hard questions inside a working day
- Scoped before it starts, reviewed every quarter
When it fits
The situations teams are usually in when they call.
- AI is going in front of customers and nobody senior owns the technical calls
- You have a CTO and want an architect beside them, not over them
- The next two quarters decide the company and the stack has to hold
- An enterprise buyer is asking questions your team cannot yet answer
A small number of teams at a time, alongside my role at TWINii, so the slice of the week is real.
One-time engagements
One problem, a fixed scope, a document at the end.
For teams that do not need someone ongoing. Each engagement is scoped on a short call, runs to a stated timeline and ends with a deliverable you can hand to your team, your board or your buyer. You know what you are getting before it starts.
The demo works and customers are next.
Demo-to-production plan
A written plan for taking the system into production: what to keep, what to replace, what to build first, with the model and vendor calls and the reasoning behind each.
- Timeline
- 2 to 3 weeks
- You get
- The plan, in writing, and a readout with your team
- I need
- The system, the design and an hour with whoever built it
One decision is holding up the quarter.
Decision session
A model, a vendor, build or buy, an agent framework. You bring the shortlist and the constraints; we work it through live and you leave with a decision you can defend.
- Timeline
- One session; the write-up within 2 days
- You get
- The decision and the reasoning, written up
- I need
- The shortlist, the constraints, the deadline
You are about to scale what you have.
Architecture review
I read the system as it runs, the code, the traces, the paths customers take, and tell you what gets expensive or fragile at ten times today’s usage and what to change first.
- Timeline
- 1 to 2 weeks
- You get
- A written review with a ranked fix list
- I need
- Read access to the code, docs and a sample of traces
Your agent is live, or about to be, and someone wants proof it holds.
Production readiness audit
Your agent tested on your real use cases against the six gates a production system has to pass: reliability, eval coverage, guardrail depth, observability, cost per resolved task and incident readiness.
- Timeline
- 3 to 4 weeks
- You get
- A scorecard, a cost baseline and a ranked 90-day roadmap
- I need
- Code, traces, real use cases and time with the owners
Releases go out when they feel ready.
Release gate
The checks a release has to clear before customers see it: the evals, the guardrails, the regression set and the rollback plan, set up with your team and left running after I step back.
- Timeline
- 2 to 3 weeks
- You get
- The gate running in your pipeline, and the team trained on it
- I need
- Your pipeline, your test data and your engineers for a few sessions
A security questionnaire is sitting unanswered.
Enterprise buyer pack
The answers your enterprise buyers will accept on security, reliability, data handling and cost, written from the system as it is, with the gaps named and a plan to close them.
- Timeline
- 1 to 2 weeks
- You get
- The answers, the gap list and the plan, in writing
- I need
- The questionnaire, the architecture and your policies as they stand
The bill is growing faster than usage.
Cost reduction pass
Where the money goes per task, by model, by step and by retry, and what to change first: routing, caching, prompt and context size, the model mix.
- Timeline
- 1 to 2 weeks
- You get
- A cost baseline and a ranked plan with expected savings
- I need
- Billing data, traces and the prompts as they run
You are about to invest in or acquire an AI company.
Technical due diligence
An independent read of what is really built: the architecture, each AI claim verified or not, the evals, the cost curve at scale, model and key-person dependency. Only where I have no stake in the outcome.
- Timeline
- About 2 weeks
- You get
- The report and a partner-level readout
- I need
- The data room, and the team for a session or two
How it starts
From first call to work in three moves.
We talk for twenty minutes
You bring the problem. I ask the questions that matter. We both know by the end whether I can help.
I read the system
What you send, under NDA if you need one: the repo, the design, the deal. I come back with a view, not a pitch.
We agree the scope in writing
What it covers, how long, what you get at the end. One page, agreed before anything starts.
Start with the call.
Twenty minutes, audio, at a time you pick. You leave with a way through the problem, or the next step towards one.
Why the call is worth twenty minutes
Judgement that has been tested in production.
years shipping
Sixteen years of production systems, the last six on AI with customers on the other end. The advice comes from incident reviews, not conference talks.
document per engagement
Every engagement ends with something your team can act on without me in the room. The reasoning outlives the call.
stake in the answer
No vendor relationships, no referral fees, no equity. Diligence only where I have no follow-on work with the target.
In their words
People who have shipped with me.
“Mandeep didn't just put in a patch, he helped redesign the system architecture to handle 5x the load without too much of extra server hardware.”
Jaspreet Singh OberoiHead of Engineering, Devnagri“Mandeep is the level-headed, methodical thinker you want on your team. He is capable of looking at the big picture, anticipating problems and planning before hand.”
Rajat HansCo-founder and CTO, Kenko AI“The guy thinks 5 steps ahead. He bridges DevOps rigor with AI workflows like it's second nature. Clean infra, predictable deploys, zero drama.”
Kamaljit SinghFounder, IndraOps TechnologiesAsked on most first calls
The questions.
Do you sign NDAs?
Yes, yours or mine, before I see anything.
We do not have evals yet. Is that a problem?
No. Most teams I meet do not. Setting the baseline is usually the first piece of work.
What access do you need?
For a review or an audit: read access to the code and architecture docs, a sample of staging or production traces and time with whoever owns the system. For a decision session: the shortlist and the constraints.
What if you find nothing serious?
Then you have a written read that says so, which is what your buyer or your board was asking for.
Do you work with teams in the US and UK?
Yes. I am in Delhi and keep hours that cover the UK and European working day and US mornings.
How is a one-time engagement different from the ongoing role?
An engagement has a fixed scope, a timeline and a document at the end. The ongoing role has a rhythm and no end date, and I carry the calls with you week to week. Many teams start with one and move to the other.
What does it cost?
It depends on the shape, and the shape is the first thing we agree. Ask on the call and you will get a straight number.