For startups and growing teams in Europe and the US
AI that ships. Not AI theater.
I’m an AI consultant and fractional Chief AI Officer. I build AI agents and agentic workflows, wire AI into the tools you already use, and help you decide which AI projects are worth doing in the first place.
Free, 30 minutes, no pitch deck. Bring one AI problem and we’ll work out what to do next.
Arpan Ghoshal
AI consultant · Fractional Chief AI Officer
- Former Director of AI Innovation at Operating Equity Partners, launching AI companies with Halsey Minor, co-founder of CNET and Salesforce
- Shipped production AI at Gisual, Hitch Works (acquired by ServiceNow), and for 10+ startups in the last year
- Deep learning research under Yann LeCun during my master’s at NYU
- EmoRoBERTa, 15M+ downloads, the open-source emotion model I created
One person. Your whole AI team.
Strategy, agents, integrations, MCP servers, RAG, controls, infrastructure. I set up everything your company needs in AI, then run and maintain it for as long as you need. No AI team to hire.
clients in the US and Europe in the last year, with AI deployed in production
Client names are kept confidential, so the consulting results below are shown without them. I’m happy to walk through any of them on a call.
Selected Work
Results from confidential client engagements, alongside named work from my roles.
AI companies taken from zero to launch
Operating Equity Partners · Director of AI Innovation
Working with Halsey Minor, co-founder of CNET and Salesforce, I led the launch of a news network built to reduce misinformation and an AI avatar legal assistant. I owned the architecture end to end: LLM pipelines, retrieval, the agent reasoning layer, evals and tracing. They still retain me to run and extend it.
Less manual workflow time
Consulting client · operations team
An operations team was copying data between tools by hand. I built a workflow that extracts and validates the data automatically and passes it to a person for review.
Lower LLM inference spend
Consulting client · production AI system
A production AI system was costing more than it needed to. I changed the model selection, routing, prompts and caching, and output quality stayed the same.
Accuracy on a first commercial model
Gisual · Machine learning engineer
I built the company’s first commercial model for detecting power outages, from geospatial data and feature engineering through deployment, then a GenAI system that spots outages on social media before conventional monitoring does.
Matching accuracy over the baseline
Hitch Works (acquired by ServiceNow) · ML engineer
I built a BERT-based model that maps skills to context for a workforce intelligence product used by enterprise talent teams, running on an AWS pipeline that crawled and processed text around the clock.
Users on a real-time AI API
Brainentech Neuroscience · ML engineer
I built and deployed a real-time audio analysis API, training the NLP models that pull 20+ engagement metrics out of speech.
Most AI projects don’t fail because of the model.
They fail because the problem was vague, the architecture got too clever, the data wasn’t ready, or nobody owned the thing from idea to production.
The first few weeks usually come down to a handful of questions:
- Where would AI make a measurable difference in this business?
- What should we build first?
- What should stay human?
- What can we ship quickly without regretting the architecture in six months?
- How will we know it works, and is it worth what it costs to run?
Three Ways to Work With Me
Each one is scoped around a result we agree on up front, and none of them are billed by the hour. Most teams start with one of these.
AI Strategy Sprint
Figure out where AI is worth it, and leave with a plan.
We go through your product and operations, rank the AI ideas by value and effort, and make the early architecture and model decisions. You end up with a roadmap your own team can build from.
Good fit if you want a clear plan before you spend money on building.
AI Build & Production Sprint
Take an idea, or a stuck prototype, to production.
I design and build it: an AI agent, an MCP server, an integration with the tools you already use, a RAG system, or an AI feature in your product. Evals, monitoring and deployment are part of the job, and it’s done when it works for your users.
Good fit if you have something to ship, or something to rescue.
Fractional AI Leadership
A part-time Chief AI Officer, for as long as you need one.
I own the AI roadmap, make the model and vendor calls, review the architecture, decide what your agents are allowed to do, and still write code when the team needs it.
Good fit if you need someone accountable for AI across quarters, without hiring a full-time executive.
Not sure which fits? We can sort that out on the call.
Anything AI? That’s me.
Strategy, agents, harnesses, integrations, MCP servers, controls and production infrastructure. The three engagements above all draw on this. Open a section for the detail.
You know AI matters. You don’t need another 60-slide strategy deck. I help you pick the opportunities worth pursuing, choose the technical approach, and keep the work tied to what the business needs. Useful when you want senior AI direction without hiring a full-time Chief AI Officer.
Fractional Chief AI Officer / AI Strategy Advisor
I help leadership teams turn “we should do something with AI” into a short, ranked list of projects. That means choosing use cases, deciding what goes first, picking the technical approach, and checking that each project pays for itself.
AI Opportunity Mapping and Roadmapping
I go through your workflows, product and internal operations to find the places where AI would save meaningful time or money. You get a ranked list and a roadmap, and you skip the experiments that were never going to pay off.
Got an AI idea sitting somewhere between “cool demo” and “how the hell do we ship this?” I can build the agent for you, designed around your problem rather than whatever framework is trending this week.
AI Agents and Agentic Workflows
I build AI agents that work across your tools and data, handling research, support, operations, sales and back-office tasks. The aim is fewer hours of manual work and fewer mistakes.
Agent Harnesses
The model is only part of an agent. The harness around it decides what the model sees, which tools it can call, how it recovers from a bad step, and when it stops. I design and build that layer: the tool loop, context and memory, permissions, retries, and the evals that tell you whether it’s improving.
Voice AI Agents
Real-time voice agents for support, intake, booking and sales calls. They respond fast enough to feel natural, can use your tools, and hand the call to a person when they should.
Most companies don’t need a new AI app. They need AI inside the product, the CRM, the helpdesk and the internal tools they already have. I do that plumbing.
AI Integration
I connect models to your data and APIs so AI shows up where your team already works: your product, your internal dashboards, Slack, your CRM or your helpdesk. You don’t have to move anyone to a new tool.
MCP Servers
Model Context Protocol (MCP) is how assistants like Claude, ChatGPT and Cursor, and your own agents, get access to tools and data. I build MCP servers for your internal systems or your product, with permissions and logging so you know what each agent did.
RAG and Enterprise Search
I build retrieval systems that let an LLM answer from your documents and internal data instead of guessing. The work covers chunking, indexing, permissions, ranking, and measuring answer quality.
AI Product Development and AI Features
I design and build AI features for your product, or new AI products from scratch. I cover the product decisions as well as the engineering, so what ships is useful to your users and more than a thin wrapper around an API.
The prototype worked. Production had other plans. Once agents take real actions and real users show up, you need rules, monitoring, evals and cost controls. This is the part I care about most, and the reason I built ctrlrun.
AI Agent Controls: Control, Observe, Analyze
Every action your agents take gets checked against your rules before it runs. Routine actions go through, sensitive ones wait for a person, and anything outside the rules is blocked. Every attempt is recorded, so you can see what each agent tried and spot when its behavior drifts. I can integrate ctrlrun, ctrl ai agents or ctrl payments into your stack, or build the same controls into your own code.
LLMOps, Evals, and AI Infrastructure
I set up what an AI product needs to run in production: orchestration, model routing, prompt versioning, eval pipelines, monitoring, guardrails, fallbacks and cost controls.
AI Architecture, Refactoring, and Reliability
I clean up AI systems that have become hard to work on: bugs, edge cases, brittle prompt chains, missing monitoring, and code nobody wants to touch.
AI Cost Optimization
I cut wasted inference spend and latency, usually through model choice, routing, prompt changes and caching, and add controls so costs stay predictable as usage grows.
Products I Build and Run
Alongside client work, I build and run tools for keeping AI agents in check. If your agents take real actions, like changing records, sending messages or moving money, I can plug these into your stack.
ctrlrun
The execution safety layer for AI agents. A Python library that checks every action against your rules before it runs: allowed actions go through, sensitive ones wait for a person, forbidden ones are blocked, and every decision leaves a receipt. Apache-2.0.
ctrl ai agents
Control, observe and analyze every action your AI agents take, including agents you can’t modify, like Claude Code, Cursor, ChatGPT or a Slack bot. Every attempt, blocked ones included, is recorded as evidence you can verify yourself.
ctrl payments
A control layer between AI agents and the money they move. Your rules decide which payments go through, larger ones wait for a person, and every attempt leaves a receipt you can hand to finance. Built on ctrlrun.
How I Work
Understand the problem
We start with the business goal, the users, the data and the budget. The model comes later.
Design the solution
I choose the architecture, models and integrations, decide how we’ll evaluate the system, and write up a delivery plan with costs.
Build and ship
I write the code and get it into production, with your team as involved as they want to be.
Measure the results
We check whether it works and whether people use it: reliability, adoption, cost, time saved and quality.
From My YouTube Channel
I talk about AI strategy and building AI products on YouTube. A few to start with:
Frequently Asked Questions
It depends on scope. A focused project, like an AI agent or a RAG system, is a fixed fee agreed before we start. Fractional CAIO work is a monthly retainer. Everything starts with a free 30-minute call so we can both see whether it’s a fit.
A focused build, like an AI agent or a RAG system, usually takes 4 to 8 weeks. Fractional CAIO engagements usually run 3 to 6 months. Some companies keep me on a retainer afterwards for advice and architecture reviews.
Yes. I clean up AI codebases, fix architecture problems, make unreliable systems dependable, add missing pieces like evals and monitoring, and help get prototypes into production.
Both. Early-stage startups usually need to get from idea to an AI-powered MVP quickly. Growth-stage companies usually want AI added to a product or operation that already exists. What they have in common is that they want results they can measure.
Yes. Most of my work is both: deciding what to build, then designing, prototyping and shipping it, and keeping it reliable afterwards.
Yes. I build agents for support, operations, sales, research and back-office work. That includes the harness around the model, the integrations with your tools and data, the evals, and the rules on what the agent is allowed to do on its own.
Every action an agent takes goes through your rules before it runs. Routine actions go through, sensitive ones wait for a person to approve, and anything outside the rules is blocked. Every attempt is recorded so you can see and analyze what your agents did. I built ctrlrun, an open-source library that does this, and I run ctrl ai agents and ctrl payments on top of it. I can integrate any of them into your stack.
Free 30-minute call
Have an AI problem? Let’s talk it through.
Bring one real challenge: an idea you’re considering, a prototype that isn’t working, a workflow you want to automate, or a production system that keeps breaking.
On the call we’ll:
- Get specific about the problem
- Pressure-test your current approach
- Find the biggest technical and product risks
- Sketch what I’d do next in your position
If I can help, I’ll explain how. If I can’t, I’ll tell you that too.
30 minutes · Remote · Free · No sales pitch
Rather write first?
Tell me what you’re working on and I’ll reply by email.