AI Agent Development
Work that eats hours of someone's week and follows rules you could write down — but never quite the same way twice.
- Agent + tool integrations
- Human approval gates
- Eval suite & run logs
- Deployment + handover
Senior-led engineering studio
From plain English to production software.
SmashGlass builds AI agents, business automations, and custom software that do real operational work — designed, reviewed, and shipped by a hands-on engineering leader.
Task from the client, in their wordsReconcile yesterday's parts orders against supplier invoices and flag any mismatch over $50.
PLANFour steps: pull orders, pull invoices, match on PO, diff line totals.
TOOLdealer_api.orders(date="2026-10-27") → 142 records
TOOLsupplier.invoices(date="2026-10-27") → 138 records
STEPMatched 136 on PO number. 6 orders and 2 invoices unmatched.
STEP4 line-total mismatches over $50. Largest $412.80 on PO-88412.
HOLDAwaiting your approval before posting adjustments to the ERP.
DONEApproved. 4 adjustments posted. Summary emailed to parts@.
Hands-on VP of EngineeringArchitecture and delivery, not account management
AI agent frameworks, shippedBuilt and deployed, not prototyped
Production systems at scaleIntegration, modernization, legacy systems
Fixed-price from $1,500Scoped up front, no open-ended retainers
What we build
Complexity in. Clarity out.
One engagement, four kinds of work. Most projects start in one of these and pull in a second.
Business Process Automation
A process held together by spreadsheets, email threads, and one person who knows how it actually works.
- Process map & failure modes
- Automated pipeline
- Exception handling
- Monitoring & alerting
Custom Software & Integrations
Two systems that should talk to each other, and a vendor who says it isn't supported.
- Integration architecture
- API clients & webhooks
- Data mapping & reconciliation
- Tests against live sandboxes
AI-Assisted Engineering
An application that still works but nobody wants to touch, and a team that can't afford to stop shipping to fix it.
- Architecture assessment
- Incremental migration plan
- Test coverage first
- Modernized, documented code
How we work
AI moves fast. Judgment decides where.
Five stages. AI-assisted implementation compresses the middle; experienced engineering judgment governs the parts where being wrong is expensive.
- 01
Discover
Understand the actual problem, the constraints, and what "done" means to you.
HUMAN-LED - 02
Architect
Decide the shape: data flow, failure modes, security boundaries, what stays manual.
HUMAN-LED - 03
Build
AI-assisted implementation against the agreed architecture, reviewed every step.
AI-ACCELERATED - 04
Validate
Tests, evals, edge cases, and a security pass before anything touches production.
AI-ACCELERATED - 05
Deliver
Deployed, documented, and handed over so your team can maintain it without me.
HUMAN-LED
Why the split matters. AI coding assistants are genuinely fast at implementation, and I use them heavily — that's how a fixed-price engagement at this size is possible at all. What they don't do is decide what should be built, where the security boundary sits, or which failure mode will take your business down at 2am. That part is the job, and it doesn't get delegated.
Capabilities
Three things I get asked to build.
Switch the scenario to watch each one run end to end.
Illustrative demonstration — not client work
Task from the client, in their wordsReconcile yesterday's parts orders against supplier invoices and flag any mismatch over $50.
PLANFour steps: pull orders, pull invoices, match on PO, diff line totals.
TOOLdealer_api.orders(date="2026-10-27") → 142 records
TOOLsupplier.invoices(date="2026-10-27") → 138 records
STEPMatched 136 on PO number. 6 orders and 2 invoices unmatched.
STEP4 line-total mismatches over $50. Largest $412.80 on PO-88412.
HOLDAwaiting your approval before posting adjustments to the ERP.
DONEApproved. 4 adjustments posted. Summary emailed to parts@.
Why SmashGlass
You get the person who does the work.
SmashGlass is a boutique studio, not an agency. There is no bench, no junior team, and no account layer between you and the engineering. You work directly with someone who has run engineering organizations and still writes and ships the code.
That matters most on the decisions that are expensive to get wrong — where the security boundary goes, which integration will quietly break under load, what should stay manual because automating it is a bad trade. Those calls come from having shipped and maintained production systems, not from a framework.
Engagements are fixed-price and scoped up front, typically $1,500–$5,000 for a first project. If it turns into something larger, we'll both know why.
- Executive-level technical judgment
- Architecture and trade-off decisions made by someone accountable for them before.
- Hands-on delivery
- The person scoping your project is the person writing it.
- Production engineering experience
- Systems that run unattended, recover from failure, and can be maintained by your team.
- AI agent architecture
- Tool design, approval gates, evals, and run observability — not a prompt in a loop.
- Reliability, testing, security
- Test coverage and a security pass are in scope, not an upsell.
- Direct collaboration
- One channel, one person, straight answers about what’s hard.
01
02
03
04
05
06
Start a project
Let's build something intelligent.
Tell me what's slow, breaking, or not getting done. You'll get a straight answer about whether it's worth building and roughly what it costs — usually within a day.