In 2020, Darnell Mitchell was two years into directing assessment services for a Southeastern university system — coordinating placement testing for roughly 8,000 incoming students annually, managing item banks across three campuses, and trying to hold together a psychometric infrastructure that had been assembled by people who'd since left. The IRT theory was sound. The item calibration process was documented. The outputs were technically defensible. But the execution was entirely manual: response data exported from the LMS, parameter estimates run in a standalone R script, cut scores applied through a spreadsheet, results uploaded back by hand. When anything broke — and something always broke — finding the error took days.
The market offered two options. Enterprise assessment platforms like Questionmark or Learnosity were technically capable but priced for testing organizations running millions of responses per year. The annual contract minimums were ten to twenty times what a mid-size university's assessment office had available. The other option was to build something in-house — which is how the R script existed in the first place. Neither option made sense. The science existed. The platforms didn't democratize it.
The founding insight
What Darnell and Isabel Reyes — who had spent seven years building IRT infrastructure at an international testing organization — identified wasn't a gap in assessment science. It was a gap in productization. The same 3PL engine that large-scale assessments relied on could be packaged into an LTI 1.3-compliant product with a calibrated API, sensible onboarding, and pricing that a department budget could absorb without a budget cycle negotiation. The question was whether you could do this while preserving genuine psychometric rigor rather than approximating it with a prettier UI around a CTT backend.
The answer was yes — but the constraints were real. A true adaptive engine required item banks with sufficient calibrated depth to select optimally at any ability level without re-exposing items. That meant building a calibration pipeline, not just a scoring engine. It meant designing the API to return standard error values alongside theta estimates — because a score without its confidence interval is a number without meaning. It meant making the behavioral signal layer (response latency, confidence patterns) a first-class output, not an afterthought appended to a completed score.
What we learned from assessment offices
The year we spent doing consulting work inside university assessment programs before building LearnVyx shaped every product decision we made. Three observations defined the spec:
LMS integration is the adoption gate, not the capability question. The most technically sophisticated assessment engine is dead on arrival if IT won't approve it. Every university we worked with had an LMS — Canvas, Blackboard, or Moodle — and any tool that required students to create a separate account or that didn't return scores to the gradebook automatically was categorically rejected by faculty. LTI 1.3 Advantage compliance (Names and Role Provisioning, Assignment and Grade Services, Deep Linking) wasn't a nice-to-have. It was the price of entry.
Accreditation language creates the buy-in argument. Institutional assessment coordinators aren't the decision-makers. Provost offices, curriculum committees, and accreditation preparation teams are. The buy-in argument that worked was "psychometrically defensible individual mastery evidence" — because regional and professional accreditors had started requiring it and institutions didn't know how to produce it. CTT aggregate scores don't answer the question; IRT individual theta estimates with confidence intervals do. Building the report exports around accreditation committee needs rather than assessment researcher needs changed the conversation.
The integrity problem was driving bad solutions. Every program we consulted with had been pressured to adopt camera-based proctoring after 2020 remote testing experiences. The resistance was significant — from learners, from disability services offices, from faculty who understood the false-positive rates. What nobody had articulated yet was that the integrity problem was fundamentally a scoring model problem, not a surveillance problem. If the score is designed to detect improbable response profiles at the model level rather than surveil the test environment, you get a more defensible integrity signal with less collateral damage to learner trust and equitable access.
Why bootstrapped, and what that means for how we build
LearnVyx has not raised external capital. That's a deliberate choice, not a constraint. The university and corporate L&D buyers we work with are not in a hurry. Procurement cycles at universities run 4–12 months. Enterprise L&D programs need pilot evidence before committing to a platform. Building a company to that pace is incompatible with growth-at-all-costs investor expectations. We'd rather spend three years building a credible measurement platform that assessment professionals trust than six months building a demo that looks impressive in a sales deck.
It also means we can be honest about what the product doesn't do. LearnVyx is an assessment engine and credentialing platform. It is not an LMS, not a content development tool, not a performance management system. Being small enough to say "we're focused on measurement quality" without a product roadmap that must justify every customer dollar raised is a genuine advantage in a market where enterprise competitors over-promise and under-deliver on psychometric validity.
We built this platform because we spent years inside the problem. That context hasn't expired.
