Agents write the code. An architect decides what ships.
ESARC is an applied-AI engineering studio for founders and small teams. A principal engineer plans the work, a crew of AI agents builds it in parallel, and nothing merges until it has been reviewed and passed its evals.
- Studio
- ESARC, applied-AI engineering
- Project
- Your product
- Drawn by
- AI agents
- Checked by
- Vaibhav Malhotra, principal engineer
- Scale
- 0→1, then up
- Revisions
- Every week
Vaibhav has shipped production systems for
- Meta
- Amazon
- McGraw Hill
- Springhouse
- MyMethod
- Spicychat
- Stuf Storage
- Scrubs Co-Pilot
Built on
- Anthropic
- OpenAI
- AWS
- Vercel
- NVIDIA
- Vapi
Why most AI projects stall
Getting a model to answer is the easy part. A product people rely on needs evals, tracing, cost limits, a plan for when the model is wrong, and one person accountable for every merge.
At ESARC that person is the architect. The agents do the typing: they draft code, tests and migrations in parallel, so you get the output of a small team with one senior engineer who knows every line that shipped.
One architect, a crew of agents
Every change goes round the same loop. The agents handle volume, the architect handles judgement, and you see the result on staging before anyone ships it.
You
1Brief
The outcome you want, who uses it, and what it must not do.
Architect
2Spec and evals
Architecture, model choices, and a written test of what "working" means.
Agent crew
3Draft in parallel
Code, tests and migrations, each on its own branch.
Architect
4Review and redline
Every diff gets read. Anything off is marked up and sent back.
Sent back: fix before merge
Agent crew
5Fix and rerun
Changes land and the eval suite runs in CI.
You
6Try it on staging
The real product, twice a week. Not a slide.
Architect
7Ship and watch
Release with tracing, cost alerts and a way to roll back.
Steps 3 to 5 repeat until the architect signs off.
How one principal + agents ships safely
The fair question is who checks the agents. The principal does, at fixed points, before anything reaches your users.
- Review gates
- Agents open pull requests. The principal reads every diff and either sends it back with notes or approves it. Nothing merges without that sign-off, and a person decides each production deploy.
- Evals and tests
- Each AI feature starts with an eval set: real examples and the outcome you expect, agreed with you before any prompt is written. Evals and tests run in CI on every change, so a regression fails a check instead of reaching a customer.
- Decisions the principal owns
- Architecture, the data model, model and vendor choices, security boundaries and anything that adds running cost. Agents can suggest options. They don’t make these calls. Each decision is written down with the reason.
- Handover
- You get the architecture notes, the decision log, the eval sets and a runbook, kept up to date as we go. A new engineer should be able to pick it up without us.
What we build
AI features rarely ship alone. They need an app around them, data under them and someone watching them in production, so we build all of it.
AI products and agents
Agents that call tools, copilots inside your app, and voice agents that answer the phone.
RAG and search
Retrieval over your documents and data, with citations and tests for the answers that matter.
Evals and observability
Eval suites in CI, tracing, and cost and latency dashboards, so a model change can’t quietly break things.
Fine-tuning pipelines
Data prep, training runs and side-by-side comparisons with the base model, for when prompting runs out.
Full-stack web
Next.js and Node apps with the APIs, auth and billing that turn a demo into a product.
Mobile
React Native and Expo apps for iOS and Android, through to the stores.
Architecture and infrastructure
Cloud setup, data models, GraphQL and queues, sized for where you’ll be next year.
Industries so far
Consumer AI, Food tech, Voice AI, Hospitality and local services, Consumer technology, AI research, Self-storage, PropTech, Entertainment, Healthcare, Education, EdTech
Selected work
What was built, for whom, and what changed. Where we don’t have a number we can stand behind yet, we say so.
- 2026Voice AI agents on Vapi, with the eval and observability work to back themMyMethodMulti-tenant voice AI SaaS
- 2026Multi-agent food planning on Pydantic AI and FastAPISpringhouseConsumer food planning
- 2025Applied AI engineering inside Meta Superintelligence LabsMeta (Superintelligence Labs)Frontier AI research lab
- 2025A pilot logbook I actually want to useClearProp (own product)General aviation / consumer
- 2025Sidney Voice AI: the storage industry's most extroverted internStuf Storage (Sidney AI)Self-storage / proptech
- 2024AI engineering consulting for an education publisherMcGraw HillEducation publishing
- 2024Structured doctor notes from unstructured audioScrubs Co-PilotDigital health / clinical workflow
- 2023A million requests a day and a shipping bill that got smallerAmazonE-commerce / Logistics
Ways to work together
Fixed-price sprints to get a first version live, then a monthly partnership once people are using it. Prices are starting points. You get a fixed quote before any work begins.
Step 1
Free 30-minute call
Free
We talk through what you have and what you want to ship, and tell you straight whether we’re the right team for it.
Step 2
Discovery week
Fixed fee, credited toward a sprint
A spec, an eval plan and a fixed quote. The fee is credited to your sprint if you go ahead.
Step 3
Pilot or MVP sprint
Fixed price, quoted after discovery
A fixed scope at a fixed price, live with real users at the end.
Step 4
Monthly retainer
Monthly, quoted per roadmap
Keep shipping every week once people are using it. Cancel with 6 weeks' notice.
Zero-to-one sprints
A fixed scope and a fixed price. You know what ships and what it costs before we start.
Paid discovery 1 week
Fixed fee, credited toward a sprint
You have an idea and need to know what it takes.
- Written spec and architecture
- Eval plan and a risk list
- A fixed quote for the build
Pilot sprint 4 weeks
Fixed price, quoted after discovery
One AI feature or workflow, proven with real users.
- One workflow in production
- Evals and tracing from day one
- Twice-weekly staging demos
0→1 MVP sprint 6 weeks
Fixed price, quoted after discovery
A whole first version: web and/or mobile, live.
- App, backend, auth and billing
- AI features with an eval suite in CI
- Runbook and 30 days of support
Build partner
Senior engineering every week, with the agent crew attached. Month to month, cancel with 6 weeks' notice.
Fractional AI lead About a day a week
Monthly, quoted per roadmap
Your team builds; you want an architect steering the AI work.
- Architecture and model choices
- Eval and review of AI pull requests
- A standing weekly call
Build partner About half-time
Monthly, quoted per roadmap
A live product that needs features shipped every week.
- Weekly roadmap and Friday demos
- Features, AI work, infra and fixes
- Evals, monitoring and cost tracking
Full-time build partner Full-time
Monthly, quoted per roadmap
You need a founding engineer now, not in four months.
- Principal engineer on your product full-time
- Agent crew running in parallel
- Hiring help when you build your own team
Is one workflow worth automating?
Put in the hours, the people and the rework, and you get the yearly value, how hard the build is, and the engagement that fits. The math doesn't use our prices, so you can take it to your own budget meeting.
What kind of AI surface?
Pick the closest workflow. Defaults shift by category, then you can tune the numbers.
Annual business case
$285,300
This is the estimated value of reclaimed time and avoidable rework. It is not ESARC pricing.
Recommended first step
Build Sprint- Hours reclaimed
- 1,794 hrs
- Opportunity cost
- $269,100
- Rework savings
- $16,200
- Implementation complexity
- Medium-high
- Confidence
- 74%
How the math works
- Manual load
- 20 hrs/week x 3 people x 50 weeks
- 3,000 hrs/year
- Effective coverage
- 65% target coverage adjusted for sensitive risk
- 60%
- Time value
- 1,794 reclaimed hrs x $150/hr
- $269,100
- Rework reduction
- 8% current rework x 45% reduction assumption
- $16,200
Why this engagement shape
The business case is concrete enough to pick one production surface and ship it behind an eval gate.
Related ESARC work
How a project runs
- 1
A free 30-minute call
We talk through what you want to ship and how we’d build it. No pitch deck, no homework. If we’re not the right fit, we’ll say so.
- 2
One week of paid discovery
You get a spec, an architecture, an eval plan and a fixed quote. The fee comes off the sprint if you go ahead.
- 3
Build in weekly cycles
Agents draft, the architect reviews, and you try the work on staging twice a week.
- 4
Launch with monitoring
Tracing, eval dashboards, cost alerts and a runbook go live with the product.
- 5
Hand over or keep going
Thirty days of support after launch. Then take it in-house, or keep building with us month to month. Stopping takes 6 weeks’ notice.
The architect
Vaibhav Malhotra, principal engineer, Vancouver, BC
Vaibhav has built production systems for Meta, Amazon and McGraw Hill, and for startups including Stuf Storage, Scrubs Co-Pilot, Springhouse and MyMethod. The work covers voice agents, RAG over clinical notes, agent memory, GraphQL backends and eval loops.
Every ESARC project has him as the one accountable engineer. He writes the spec, reviews what the agents produce, and is the person you talk to. He’s also a working pilot, which is why ESARC’s own product is a pilot logbook.
Got something that needs to ship?
- Yes
Book a call and bring the messy version. We’ll tell you what the first sprint would cover.
- Not yet
Email vaibhav@esarc.dev with where you’re stuck. You’ll get a straight answer, not a pitch.