HV-LABS. — An AI practice by Husien Vora.

Five scenes I keep walking into.

Not pains. Decisions that weren't made.

The Bill.

The bill arrived. One line. No breakdown.

The team had to guess which feature spent what.

The guess was wrong.
The New Hire.

A new engineer joined. They wanted to try Claude.

They opened the codebase. Searched openai. Twenty-three hits.

They went back to GPT-4.
The Smart Model.

The smart model worked. They used it for everything.

The bill grew quietly.

Half the calls were paying premium to do the work of a switch statement.
The Agent.

The deck said "AI agent." The code said for-loop.

The Series A didn't ask. The Series B did.

Now the founder cares.
The Bug.

A bug report came in. "The AI gave a wrong answer."

The engineer shrugged. The prompt had been edited three times since the last test.

There was nothing to debug against.
The Team.

Support typed the same answer twelve times a day. Sales wrote the same outbound email all afternoon. Engineering reviewed the same migration patterns every sprint.

The product had AI on it.

The team didn't have AI inside it.

None of these are bugs. They're decisions that nobody made. The audit names them.

What changed.

There are two halves to this now.

One: the right call.

Cheap models keep getting cheaper. Smart models keep getting more expensive. Every prompt is a budget decision. Half the calls in a typical stack should be a smaller model. A quarter shouldn't be a call at all.

Two: the right place.

Most teams have AI on the product. They don't have AI inside the team's actual work. Engineering loops. Support drafts. Research. Ops. The team that bakes AI into its own workflow ships twice as much as the one that bolts AI onto its product.

Picking the right model saves money. Putting AI in the right place compounds. Most stacks have done neither sort.

The build.

Only if you want it.

The "should this be AI?" pass. Feature by feature. Honest answers.
Workflow integration. AI inside engineering, support, research, ops. Where it compresses the work, not where it decorates the product.
Eval frameworks. So you know when a prompt drifts before users do.
Model abstraction. Swap providers in a day, not two weeks.
Cost observability. Per feature. Per user. Per prompt.
Prompt versioning. One source of truth. Diff-able. Reviewable.
Real agents. If you actually need them. State, tools, planning, recovery.
New capabilities. Whatever shipped this year you should already be using.

Fixed-scope or hourly. I start with one piece. Prove it. Then we scope the next.

The advisory.

After the build. Or instead.

Monthly checkpoint. Cost discipline, eval drift, model choices, what shipped this month that should change yours.
Architecture advisor. On call for the calls worth thinking about. Vendor changes, new capability adoption, agent design, hire-or-buy.
Embedded developer slot. Hands on keyboard when you need a specific feature shipped or rebuilt. Same person, same context, no re-onboarding.
Team enablement. Helping your existing engineers level up on prompting, evals, and agentic patterns instead of doing it all for them.

Retainer or hourly. Scale up, scale down, pause, end. The audit pays the diagnostic. The advisory makes sure it stays useful.

About me.

It's just me. On purpose.

I'm Husien Vora. I built this as a one-person practice because the work doesn't need a team yet, and pretending it does would be the first thing I'd tell a client not to do with their own stack.

AI work: building with the OpenAI API since it shipped publicly. LangChain and agent frameworks since the earliest usable versions. Multiple production agent runtimes shipped (Eliza, Virtuals, OpenAI Tools, MCP adapters). Claude Code and Cursor in daily rotation, used to compress real engineering work, not to perform it. Involved in research discussions past and ongoing with researchers at Stanford, Meta, and Indian AI startups.

Engineering background: four years of production work. Real-time trading systems at HFT-grade latency, mobile applications past 100K downloads, distributed protocols built from scratch. The plumbing experience pays back when AI infrastructure has to actually work at scale.

The same conversation kept happening with founders I respect. "Is my AI any good? Am I burning money? Should I rebuild?" The honest answer was usually yes, no, and not yet. I'd rather give that answer in a structured way, fix what's worth fixing, and stick around to help the workflow keep up with what ships next.

github.com/Husienvora · linkedin.com/in/husien-vora · husien.s.vora@gmail.com

AI running unnoticed is a hire. One you pay to mess things up.

Not pains. Decisions that weren't made.

The Bill.

The New Hire.

The Smart Model.

The Agent.

The Bug.

The Team.

There are two halves to this now.

Free. Really.

Only if you want it.

After the build. Or instead.

It's just me. On purpose.

The first call is free. The diagnostic is yours either way.