Platforms & Verticals
Learning platforms built around how your programme actually runs — assessments, cohorts, credentials and analytics — rather than around what an off-the-shelf LMS assumed.
Most organisations should not build a learning management system. If your requirement is hosting courses, tracking completion and issuing certificates, a mature product will do it better and cheaper than anything custom, and we will tell you so on the first call. Building is the right answer in a narrower set of cases: when the learning experience is your product rather than an internal function, when assessment or credentialing logic is genuinely specific to your domain, when you need learner data flowing into systems no LMS integrates with, or when per-seat licensing across tens of thousands of learners has made the buy option absurd. Those cases are real and they are where we work. ZenMagix builds learning platforms for edtech companies, training providers, certification bodies and enterprise learning functions in India and abroad. The engineering that matters is rarely the video player. It is assessment integrity, progress modelling that survives a learner switching devices mid-module, standards compliance so content is portable rather than trapped, and analytics that tell an instructor which learners are quietly falling behind before they drop out entirely.
Buy when learning is a function you operate. Build when learning is the product you sell, when the assessment logic is genuinely domain-specific, or when per-seat licensing at your scale has stopped making sense.
The buy case is strong and we make it honestly. Moodle, Canvas, TalentLMS and their peers have absorbed a decade of edge cases in enrolment, reporting, accessibility and compliance that a new build has to rediscover at your expense. If your requirement is hosting content, tracking completion and issuing certificates, buying is correct and the conversation should be about implementation rather than development. The build case turns on four conditions. First, when the platform is the product — an edtech company's differentiation lives in the learning experience, and shipping a lightly skinned commodity LMS means competing on price alone.
Second, when assessment or credentialing is domain-specific: adaptive testing with item response theory, practical skill assessment with structured rubrics, regulator-defined competency frameworks, multi-stage certification with revocation rules. These are things off-the-shelf platforms handle badly or not at all. Third, integration depth: when learner records must move bidirectionally with an HRIS, a student information system, a CRM and a billing platform, integration work frequently exceeds what the platform itself would cost.
Fourth, unit economics — per-seat pricing across fifty thousand learners can make a custom build pay back within a year, and that arithmetic is worth doing properly before committing to either path. Where the answer is mixed, we often recommend a hybrid: an existing LMS for content delivery with custom assessment and analytics layers around it.
Content delivery is a solved problem. Assessment integrity, item quality, adaptive sequencing and defensible scoring are where a platform earns or loses credibility.
Learners forgive an unremarkable video player. They do not forgive an assessment that marks a correct answer wrong, loses a submission, or produces a score nobody can explain. So this is where the engineering effort belongs. Item banking comes first: questions as versioned objects with metadata for topic, difficulty, discrimination and exposure, so the same test can be generated fresh per learner without repeating items or leaking an answer key across a cohort.
Adaptive sequencing built on item response theory produces shorter, more precise assessments than fixed-form tests, but it requires calibrated items and enough response data to calibrate them — we are explicit about that prerequisite rather than promising adaptivity on day one. Integrity measures scale with stakes. For low-stakes formative quizzes, randomised item selection is sufficient. For certification that carries professional weight, you need question randomisation, time controls, lockdown or proctoring, and anomaly detection on response patterns — an implausibly fast sequence of correct answers on hard items is a signal worth flagging.
Proctoring has real privacy implications, particularly under India's DPDP Act, and we implement it with explicit consent, minimal retention and a documented lawful basis rather than recording everything by default. Scoring must be defensible. When a learner disputes a result, you need the item version they saw, their responses, the scoring rule applied and the timestamp — reconstructable in minutes. Platforms that cannot do this end up settling disputes on goodwill.
Strong at practice generation, formative feedback and instructor triage. Weak, and risky, at final grading and at answering questions outside the course material.
AI tutoring is the most requested feature and the one most often specified badly. It works well in a few specific places. Practice generation: producing additional exercises at a chosen difficulty from an existing item bank, which addresses the real problem that learners exhaust the available practice. Formative feedback on open responses, where the value is a fast explanation of why an answer is incomplete rather than a grade. Socratic tutoring constrained strictly to course material, refusing to answer outside it — the constraint is the whole design, because a tutor confidently teaching something the course does not cover undermines the instructor.
And instructor triage: surfacing which learners are struggling and on which concepts, which is genuinely hard to see manually across a large cohort. Where it does not belong is summative grading of anything consequential. Model scoring of essays correlates reasonably with human grading in aggregate and fails unpredictably on individual cases — the unusual but excellent answer, the response in a dialect the model handles poorly, the argument that is correct but unconventional.
For a certification that affects someone's employment, that variance is not acceptable, and a learner appealing an AI grade deserves a human decision. Our default is AI-assisted human grading: the model drafts feedback and flags likely scores, a human decides. It saves most of the time without moving the accountability.
SCORM and xAPI portability, WCAG accessibility and DPDP-aligned data handling are what enterprise and institutional buyers check. Failing any of them stops a deal regardless of product quality.
Platforms that sell well are the ones that clear procurement, and procurement checks a predictable list. Content portability comes first: SCORM 1.2 and 2004 for compatibility with existing corporate libraries, and xAPI or cmi5 for the granular activity data modern learning analytics need. Buyers ask because they have been trapped before, and being able to demonstrate export of both content and learner records answers the objection immediately. Accessibility is not optional.
WCAG 2.2 AA is the practical benchmark: keyboard navigation throughout, screen reader support in the player and the assessment engine, captions on all video, sufficient contrast, and no interaction that requires precise timing. Public sector and large enterprise buyers in most markets ask for a conformance statement, and retrofitting accessibility into a finished platform costs several times what building it in does. Data protection under India's DPDP Act requires explicit consent for the personal data you collect, purpose limitation, defined retention, and a mechanism for a learner to obtain or erase their record.
Where you serve EU learners, GDPR adds its own obligations, and where minors are involved both frameworks impose stricter requirements. Proctoring data is the sharpest edge here — video of a learner's room is sensitive, and collecting it without a documented basis and a short retention window is a liability rather than a feature. We build these in from the start because each one is far cheaper as a design decision than as a remediation.
What you get
An honest analysis of whether you should build at all, including per-seat cost modelling at your real learner numbers, a review of whether an existing platform meets your assessment requirements, and a hybrid option where that is the better answer.
Course authoring or import, cohort and enrolment management, progress tracking that survives device switching and connectivity loss, credential issuance with verification, and a learner experience designed for your actual programme structure.
Versioned item banks with topic, difficulty and discrimination metadata, exposure control, randomised form generation, optional adaptive sequencing, integrity measures scaled to stakes, and fully reconstructable scoring for disputed results.
Practice generation from the item bank, formative feedback on open responses, tutoring constrained strictly to course material, and an instructor view surfacing which learners are struggling on which concepts before they disengage.
SCORM 1.2 and 2004 support, xAPI or cmi5 activity streams to your learning record store, WCAG 2.2 AA conformance with a documented statement, and DPDP-aligned consent, retention and learner data access mechanisms.
How we work
Two to three weeks reviewing your requirements against existing platforms, modelling per-seat costs at real learner volume, and testing whether your assessment logic can be expressed in a bought system. A recommendation to buy is a common and entirely successful outcome.
Learners, cohorts, courses, competencies, assessments, attempts, credentials and the rules connecting them. Getting this wrong is expensive to correct later because every report and every integration depends on it, so we spend real time here.
Item banking, form generation, attempt handling, scoring and reconstructability before anything else, because this is what learners judge the platform on. Content delivery is comparatively straightforward and comes after.
Six to eight weeks with an actual cohort on the platform, instrumented for drop-off, assessment disputes and support contacts. Standards compliance, accessibility audit and load testing complete before broader rollout.
Every phase ends at a decision point you can stop at — see how that works across fixed-scope projects, embedded pods and retainers.
Stack
Questions
Use an existing one unless learning is your product, your assessment logic is genuinely domain-specific, your integration requirements are deep, or per-seat licensing at your learner count has become absurd. We run that analysis as a paid two to three week piece and frequently conclude that you should buy.
A focused platform with real assessment capability takes four to seven months to a piloted first release. Anything quoted at six weeks is a content player with a completion tracker, which may be all you need — but it is not what people mean by an LMS.
Yes, SCORM 1.2 and 2004 both. We also emit xAPI or cmi5 for granular activity data, and we make export of both content and learner records demonstrable, because the buyers who ask about standards are asking because they have been locked in before.
For formative feedback, yes, and it is genuinely useful. For summative grading that affects certification or employment, our default is AI-assisted human grading — the model drafts feedback and flags a likely score, a person decides. Model scoring fails unpredictably on unusual but excellent answers, and a learner appealing a grade deserves a human.
Yes, with integrity measures scaled to the stakes and privacy treated seriously. Proctoring video is sensitive personal data under the DPDP Act, so we implement explicit consent, minimal capture, short retention and a documented lawful basis rather than recording everything by default.
We build to WCAG 2.2 AA across the player and the assessment engine, with automated checks running in CI and a documented conformance statement for procurement. Retrofitting accessibility costs several times what building it in does, so it is not treated as a later phase.
Questions about cost, timelines, IP ownership and data residency are answered on the general FAQ, and how this practice came out of blockchain infrastructure explains why we build the way we do.
Related
Inspection, counting and tracking systems that work in real warehouse lighting.
ROS, SLAM, sensor fusion and fleet management. Software, not hardware.
AI-powered web applications: streaming interfaces, in-app retrieval, cost control and evals in CI.
See also: take your course library multilingual with video localization and dubbing · extend the platform into your wider systems with custom software development · ground the tutoring layer properly with RAG and LLM integration
Tell us what you are trying to build. We will tell you honestly whether we are the right team for it, and what it would realistically take.
