Track record
Four engagements in putting AI to real work — three delivered for clients, one we run our own business on. Different problems, one standard: the result had to be trustworthy in production, not just impressive in a demo.
Engagement 01
Role: Subcontractor to a consulting partner · Healthcare services client
Brought onto a consulting engagement to handle a specific technical problem: a healthcare services provider — newly formed by merger — needed a simple tool for its customer service agents to look up which payer was responsible for a given order. Simple to use, hard to build. The rules that governed those decisions lived in roughly 140 Excel spreadsheets, and after the merger, no two were structured quite the same way.
The tool wasn't the hard part. The translation was. Every one of those 140 spreadsheets had to map cleanly onto a single Snowflake schema — and the data had to be trustworthy once it landed, or the tool would be worse than useless.
We used AI to do that work: read the target schema, map each spreadsheet's quirks onto it, load the records, and validate every row on the way in. When a row didn't hold up — a malformed code, a missing routing target, a placeholder where real data should be — it was flagged and set aside for a person to review, not quietly imported. Bad data never made it into the system.
The load ran in about eight hours and moved more than 27,000 records, with roughly 2,200 inconsistent rows caught and held back for human review. By hand, the same work would have taken four to five days — and would almost certainly have carried errors through undetected. The agents got a tool they can trust, built on a clean, auditable foundation instead of 140 spreadsheets no one could reconcile.
Engagement 02
Role: Subcontractor to a prime · Federal defense client
A prime was building an AI knowledge base to help fleet maintenance technicians get answers faster. The raw material was a backlog of real maintenance tickets, exactly the kind of operational history that makes a knowledge base useful. There was one problem standing between that data and the system: the tickets were full of things that must never land in an AI corpus. Vessel identifiers, personnel names, operational movements. Doing it by hand was slow and error-prone. Running it through AI to clean it was the obvious shortcut, and the obvious mistake, because that would mean feeding the exact sensitive data into the exact place it can't go.
So we split the problem. We used AI to build the sanitizer, fast, but built it to run with no AI in the pipeline, so the sensitive data itself never touched a model. The tool ran entirely in-boundary, with nothing leaving the client's environment. It replaced sensitive values with consistent stand-in tokens so the records stayed coherent and useful rather than blacked out, left the technical content that gives the tickets their value untouched, and flagged anything it wasn't sure about for a person to review instead of guessing. Every original was preserved; a human signed off on the output before any of it moved.
That let the prime move forward with a corpus that was safe to build on, and get there ready, without having exposed a single sensitive detail to do it. Sometimes deploying AI properly means being deliberate about where it belongs and where it doesn't. Drawing that line correctly, and building the thing that respects it, is the same judgment we bring to the work AI should do.
Engagement 03
Role: Subcontractor to a prime · Federal defense client
A prime's client served the Department of Defense with a large, established application. Officials used it to look up Congressionally approved projects, check current progress with full drill-down detail, and submit updates and requests against a given project or budget. The system worked. The problem was that people weren't using it. Each service branch had its own slightly different, uniformly complex forms, and the friction was killing adoption.
The obvious fix, rebuilding the interface, was the wrong one. It would have meant disturbing a working, sensitive system and absorbing all the risk that comes with that. So we didn't touch it.
Instead we built a separate, FedRAMP-aligned companion application that reached the existing system through its own APIs using MCP, adding capability alongside the original without modifying a line of it. That companion layered on three things the base system couldn't offer: AI-powered contextual search across the projects, AI-generated insights that surfaced where applications had stalled or been abandoned and trended those patterns over time, and a guided chatbot that walked a user through the process based on their specific branch of service.
The original application kept running exactly as before: same data, same system of record, zero added risk to it. What changed was the experience around it. The parts that were driving people away got easier, and the branch-specific complexity that hurt adoption became something the AI handled for the user instead of something the user had to fight through.
Engagement 04
Role: VCG's own product
ReqStaff runs the actual work of a consulting and staffing business, with AI doing real work at every stage. It sources and screens talent, parsing and reformatting resumes and scoring technical screens. It drives the sales and recruiting pipelines. It tracks time in the field and rolls it up into operational and profitability reporting. This isn't AI bolted onto a CRM as a feature. It's AI carrying the workload across the whole operation, from first contact to billable hour.
The cherry on top is the sales assistant: a multi-tool AI that works the way a good salesperson wishes they could. Ask it a question and it queries live CRM data, researches companies and the labor market on the web, reads an uploaded RFP or job description, drafts a lead, and remembers every prior conversation through semantic search, so context is never lost between sessions. It doesn't just answer, it acts, across a whole toolkit, on your real data.
And it runs under discipline. Every AI feature operates within a governance layer, with each model call registered, assigned, and audited, and sensitive personal data deliberately kept out of the AI's memory. It's proof we build AI in as the foundation of a production system real people depend on daily, and that we govern it properly while we do.
Tell us what you're working on and we'll tell you honestly whether we can help.
Get in touch →