We build production AI systems, and measure them before we ship them.
Dedicated platforms, agent systems, and the evaluation infrastructure that tells you whether either one can be trusted. We operate what we build, and we run it on our own businesses first.
What we build
Systems that run a business, not demos that describe one.
Every engagement ships working software into production, with a human approval gate where a mistake costs money, and kill switches at every level where it can.
A private AI system per customer
Own database, own pipeline tuned to your inventory or workflow, your brand, your domain. Built against a written specification and delivered with supervised go-live.
Voice, phone, and workflow agents
Agents that take calls, qualify leads, book meetings, and write to your system of record, with the logging and audit trail to see what they did.
Know how much your judge disagrees with itself
Retest variance, verdict stability, and noise-aware comparison for LLM judges, reviewers, and graders, before a number from them is allowed to drive a decision.
Full-stack, founder to first customer
Marketplaces, intake apps, CRMs, and web products for non-technical founders, built end to end, then handed over or operated.
Technical leadership on contract
Architecture, vendor and model selection, hiring bar, cost control, and the shipping cadence, for teams that need a CTO's judgment before they need a CTO's salary.
Research that ends in running code
Working papers with reproducible models, including defense-technology R&D. Scoped and discussed directly.
In production
What we have shipped and run.
Each of these is live. Several are operated by us or by customers on dedicated infrastructure.
| System | What it does | Class |
|---|---|---|
| roseglass.app | Dedicated intake-to-listing platform for IT asset disposition companies: photograph, identify, price, approve, publish. | Dedicated platform |
| supply.exchange | Exchange where surplus producers post disposition requests and certified ITADs bid in open threads. | Marketplace |
| bullpen.sale | AI sales agents that answer the main line, call new leads within a minute, qualify, book, and log to Salesforce. | Agent system |
| gitdr.ai | GitDoctor: read-only inspection of scheduled loops, queues, keys, and model spend, with a reversible fix for each finding. | Developer tooling |
| indbyagent.com | One invitation link that renders as a party page for people and as structured data agents can RSVP to. | Agent-native web |
| noisefloor-llm | Open-source retest-variance measurement for LLM judges. MIT, no dependencies. | Open source |
Evidence
Numbers we can show, and what we measured about our own tools.
First customer result, then what happens when we point our measurement tool at things, including our own.
BST (Bay Surplus Technology), first live ITAD on the platform, August 2026. Results are one customer's one month; we publish them as that.
Measured: an LLM reviewer
An acceptance verifier from an open agent framework, judging labelled code patches ten times each. Correct on every unambiguous item by majority. Not stable.
items 10 (+3 easy)
judge Sonnet 4.6
correct 8 / 8 unambiguous
stable 7 / 10
pairwise flip 6.0% on hard set
<2.2% on easy set
slow draw = the flip, on both
models tested
Measured: our own instrument
A four-model agreement gate we built. It never opened, and we assumed noise. Retest said otherwise, and we published the correction.
cells tested 25
disagreement > noise 23 / 25
lens spread 50-1000x
its own retest
cause reproducible
disagreement,
not jitter
Full write-ups, logs, and the tool that produced them are in the open-source repository. We record predictions before running and note where we were wrong.
How we work
Four rules we do not bend.
- Measure the judge before trusting the score. Any model that grades, reviews, or approves gets a retest before its output drives a decision.
- Human gate where money moves. Approval queues, role-based access, and kill switches at platform, tenant, and feature level.
- Dedicated infrastructure. One customer, one database, one credential vault. No shared tenancy to explain to your auditor.
- Say what failed. Findings that contradict our own claims go in the write-up, not in a drawer.
Engagements
Three ways to work with us.
Fixed-price where the scope is a written specification; time and materials where it is not.
Fixed-price platform build
Specification, build, on-site onboarding, supervised go-live. A year of platform use and no-charge debugging, then a flat monthly platform fee.
Stand-in CTO
A senior engineer owning architecture and delivery on retainer. Founders get a technical counterpart; teams get a second opinion with an executable answer.
Venture-studio build
For non-technical founders working with the RoseGlass studio: we are the engineering arm that takes the product from idea to a system with customers.
Contact
Tell us what you want running.
Two sentences on the problem and what it costs you today is enough to start.
office@roseglass.dev