We help medium-sized businesses become AI native and AI run.
The most tedious task white collar employees do becomes a lot easier and a lot faster. Rote operations work gets minimized, and most importantly that unlocks the team to do more important high value work, because now they have clean data, fast access to it, and a learning machine that can make them extremely competitive in the marketplace.
Readiness map: ten business days, fee credited against the build.
Your data lives on other people's computers.
The last 20 years of software was SaaS (software as a service), that means storing all of your business data on somebody else’s computer, servers run by special companies like CRMs (customer relationship management systems), payroll systems, internal wikis, ticketing systems, ERPs (accounting and operations systems), code storage. The list goes on. Now to use AI you need to consolidate the data, because it needs to understand permission, who has access to what temporarily or permanently, it needs to understand context because even some small fraction of missing context will lead to the job getting done incorrectly, and this is typically what humans did in the pre-LLM era, before large language models (LLMs) like ChatGPT and Claude.
So the job now is to consolidate your data, make it available via MCP (Model Context Protocol, the standard way AI tools plug into your systems), create MCP gateways, create access controls, and here’s the hardest part, consolidating down across vendors in a central place like a database or a data warehouse. There’s no linkages between data. There’s no relationships between data, so you may have a person record in multiple different places and it might be phrased differently, or may have a different email address, and for that you need the evals, automated tests that check the answers against ones you already know are right. The evals will make sure that your data pipeline is properly joining your records.
It’s not a one time job. You have to set it up and monitor it and tweak it, improve it over time as your data changes, and it’s different for every business. An eval suite and an LLM prompt for one law firm or one financial services firm, it’s going to perform differently than for another.
Your AI rebuilds the same answer every time you ask. Joining data live is brittle, slow and expensive.
The current status quo is to use MCPs from all your vendors, connect them to your Claude or OpenAI accounts, and the agents and the LLMs end up joining the data on demand in real time. And it works. It’s working for you today. The problem is that it’s brittle, doesn’t always work, and wastes employee time as they wait for the agent to do repetitive work each time.
Let’s say you ask who is our best performing partner or salesperson. It has to go into at least two or three different systems to understand customer retention, customer acquisition cost, sales cycle, and join that data in real time using very expensive tokens, which is how AI providers bill. If that question comes up often, and it does, it’s much more efficient to pre-join the data in your database and deliver that in any form, whether it’s a dashboard, an LLM conversation or an internal report. That artifact generation becomes free. So it’s a trade-off between doing things upfront or doing things on demand. Doing things on demand works today, but it doesn’t scale and it’s very expensive.
Hiring a senior staff engineer to handle this for you is an expensive affair. It’ll cost you at least $400,000 to $500,000 per year all in, and you’ll have the same result, but it comes at a high cost. Without one there are many questions the firm can’t answer now, not without staffing a data scientist for at least four hours to spin up a custom script or a notebook.
Fivetran and Airbyte are excellent pieces of software that excel at moving data in a deterministic way. What they don’t provide is a way for you to fix your data and to enrich it properly. They don’t know your business. Only you know your business.
Organized data becomes your edge.
Most companies I talk to have existing systems with imperfect data. You combine these two situations, where you have distributed data and it’s not groomed, so it’s inaccurate or poorly formatted, and that amplifies the problem for an LLM that you’re talking to today. It has to clean up data, cross reference it and join it. But that’s not the ultimate reason why you want to do the upfront work to consolidate and organize your data.
There are two other primary reasons. One is you can build proprietary edges in your business, or advantages in your business, based on clean organized data. So if you’re a law firm looking to organize your client contact to understand what is the next best step for each client, you can build a system that can prioritize client touch points.
If you have all the context correctly put together you can even build proprietary models in the future, once costs come down more significantly. Doing it on demand doesn’t set you up well for a future where the LLMs and the agents do real work. And the benefit of consolidating your data is that questions become much faster and easier to ask.
You need to assess where you are and where you want to go before building anything.
There are lots of steps along the way to building and consolidating your data in a centralized place, but first you need to understand where you're starting from.
Readiness map
The map helps uncover opaque parts of their business that maybe no single person knows about. It helps them identify gaps in their systems that AI can solve and AI cannot solve. It's almost a mandatory scoping exercise, because it prevents critical surprises from increasing the bill later on.
- Systems of record and their owners
- Join map: where your systems disagree about the same client, with record counts
- Data-quality counts
- Export rights, go or no-go per source
- Three AI uses reachable in 90 days
- Fixed price for the build
Foundation sprint
The most tricky, challenging part, that requires tons of customization for each individual company, is the data matching or the data cleanup or the data grooming, whatever we want to call it. If there's additional data sources that we find out about later, that increases the scope of the project. Sometimes there's even vendor negotiations that have to take place to extract the data. So it's really important to know where your data is and who owns it and how to get it.
- One store in your cloud account
- Two to four sources
- Gateway: sign-in, approved tools list, a log of every request
- Two or three tools
- Test questions with known answers, per tool
- Operating manual, data documentation, passwords in your own vault
Run retainer
Data is not a static thing. More new data comes into your world, and the LLM prompts, the models, they have to change as new data comes in or as new capabilities come out. As costs come down, there's optimizations you can do to cut down operating costs. It typically doesn't require a full-time engineer for most medium-sized businesses, but it's a non zero maintenance cost where you need an expert who knows the data and knows the industry to come in, calibrate it and tune it as time goes on.
- Data feeds monitored and repaired
- Monthly answer-quality report
- Monthly cost report
- New tool or source each quarter, middle tier up
- Extra hours agreed in writing
- No data is better than wrong data. Unmatched records go to a review queue.
- 24 months of history per source, older history quoted separately
- Vendor data access and fees are yours, and waiting on vendors moves the schedule
- Fixing records in your source systems stays with you
- Extra tools quoted separately
- Half on signature, 40% at the midpoint, 10% on an acceptance test
Going directly to a build can be very, very expensive, if you get it wrong.
Ongoing services
Data readiness
Clean up SharePoint and Google Drive permissions and file structure. Every AI rollout breaks on this first.
Seat provisioning and governance
Choose the approved tools, set up single sign-on (SSO), data loss prevention (DLP) so client data doesn’t end up in the wrong tool, and usage policies, and find shadow AI: tools staff use without approval.
Managed workflows and agents
Intake, accounts payable, scheduling, document drafting, inbound call handling. A monthly fee to run, monitor, and fix them as models and vendors change.
AI-aware security
Prompt injection, where hidden instructions in a document or email hijack an AI tool. Deepfake and business email compromise (BEC) defense, against someone impersonating a partner or vendor to redirect a payment. Data leakage.
Decision quality
Trusting AI systems with real world responsibilities requires measuring performance. How we measure it.
Why not just…
…wait for your software vendor to ship this?
The vendor's interest is having you access the data you have in their database, their cloud, their computer more efficiently, so they can potentially upsell you. But the problem is that most modern businesses don't have all their data in one place. If you go with every vendor's AI solution, you're going to have a half dozen to two dozen different AI solutions that don't talk amongst each other. The secondary reason is that you really want to take control of your data and minimize leakage of data, because as AI becomes more capable you can imagine the data that your business inputs and outputs can be used to steal your business. If they do eventually start competing with you, you want to maintain your advantage, and that's your data.
…get an agency to do it?
With an agency, what you lose is the domain knowledge that is needed to do a good job. What you gain with us is that the person who built your whole system soup to nuts knows your data, eventually as well as you do or maybe even better. With an agency you get a revolving door of people, and as time goes on, you may see a gap in the level of quality they output. Their interest is typically billing by the hour. We are interested in outcomes, not just billable hours.
Evals: checking AI’s work the way you’d check a new hire’s
AI can give a different answer to the same question twice. Evals are a way to increase and maintain reliability. The ultimate answer is to measure the performance of the system via evals. Enriching data, joining data, having data accessible by the right people at the right time: those are all metrics that can be calculated and shown on a continuous basis.
The evals run continuously, some daily, some weekly. They’re checked by different stakeholders. The first line of attack to get evals in a good place is using a high-quality model in the first place to collect true ground truth data, the answers you know are correct. After that it really depends on the context. If it’s sales data, the sales team should handle or support the evals and the ground truth. If it’s marketing, marketing needs to take care of their side of the house. I build the tools to make that happen.
Heads of IT demand observability, being able to see what the system did. It’s important to know where data goes to and how. Most firms have data obligations of some sort, and nobody wants to be in the headlines for a data breach.
Clean data does increase model performance. It unlocks real work. This is a must for all businesses in the new post-AI world.
Who
Ilya is a seasoned professional engineer and product manager with over fifteen years of experience in tech.
What’s most important is my recent experience working in AI as an entrepreneur building my own products. I’ve built ContentFlow, which is a podcasting intelligence platform, and Acute Markets, which is a mini Bloomberg type terminal for stock, bond and macroeconomic data.
At a venture firm I’ve done exactly this. When I joined, the first thing I did was create an MCP gateway, then drove adoption for it across a wide array of MCP clients including Claude, ChatGPT and Genspark. The next logical step after that is, how does the investment team benefit from this? Really the goal is to consolidate the data, cause you can imagine a really basic task such as write an email needs a ton of context. It needs context from the meeting transcript, from the email thread, from whatever notes were entered on the internal CRM for the person being contacted. That’s already three or four systems you need to pull data from in order to just write a very short two or three sentence email.
- Senior technical program manager, Cruise
- Senior engineer, Voyager, 2018 to 2022
- Founder, Celery, until 2017
- Client coverage, Goldman Sachs REDI
- Fixed income pricing, Thomson Reuters
- Runs in your cloud account, under a role you can revoke
- You own the data, how it's organized, the tools and the test questions
- No training on your data
- Credentials in your vault
- Export and retention rights checked per source
- Fixed price, scope in the contract
Security and architecture questions: ilya@megalabs.co
Find out what your data is actually like before you budget for it.
I think it behooves every business today to examine where their data is and try to consolidate it in order for AI to do real work. I would ask what their current posture is with AI, with LLMs, and how they're using them, and what their alternative is. Ten business days, and the fee comes off the build if you book one within thirty days.
- Days one to fourSystems in, matches checked
Each system of record walked through, then a check of where your systems disagree about your top 500 clients, projects or matters
- Days five to tenVerdict, uses, price
Go or no-go per source on export rights, then a fixed price naming sources, records and tools
- Read access only, or a guided screen share
- Nothing changed in any source system
- Only counts appear in the document
- No training or benchmarking on your data
- Mutual nondisclosure agreement (NDA) before kickoff
- You keep the document, even if it says not to build