CHALLENGE · 01
Personalization at scale
Every learner has a unique knowledge graph. Static curriculum can't branch fast enough to serve all of them.
IRT + Knowledge graphIndustry · 02 of 06 · K-12 · Higher Ed · L&D
Learning that adapts to the learner — at the speed of a single keystroke.
Classrooms have one teacher and thirty students moving at thirty different paces. We build the platforms that close that gap — adaptive paths, AI tutors that respect the syllabus, and the analytics that tell a teacher who needs help before the parent-teacher meeting does.
+34%Course-completion lift
12×Cheaper formative grading
−9 dAt-risk detection lead-time
learner · maya.s · grade 9PATH · ALGEBRA II
Linear equations
12 / 12 mastered
Quadratic functions
7 / 12 · in progress
Systems & matrices
recommended next
Polynomial graphs
unlocks at 80%
Exponentials & logs
unlocks at 80%
AI tutor · session 14
Maya — you keep slipping the sign on negative discriminants. Want to try 3 worked examples, or jump to a short video?
The reality on the ground
Why the 30-student class is a 1960s constraint.
The cost of one teacher is finite. The cost of one tutor that runs all night is not. The question isn't whether AI belongs in the classroom — it's how it shows up without removing the teacher from the centre of it.
FACT · 011 → 30
Classroom ratio that hasn't moved in 60 years. Bloom's 2-sigma problem has been waiting for technology to solve it — and now it can.
FACT · 0274%
Of students fall behind on a single mis-mastered prerequisite. By the time the gap shows on a test, three weeks of catch-up is required.
FACT · 03$8B
Global LMS market. Yet most institutions still have an LMS that's essentially a 1998 file-share with a forum bolted on.
FACT · 0448%
Of higher-ed dropout decisions are made in week 6. Engagement signals to predict it exist by week 3 — if you're looking.
The hard problems
What keeps a Chief Academic Officer awake at 02:00.
Every educator we talk to says the same thing: "I have the data, I just can't see it in time." That's the problem we keep solving.
CHALLENGE · 02
AI tutors that don't hallucinate
A generic LLM will invent a prerequisite, contradict the textbook, or confidently solve the problem the wrong way.
Curriculum RAGCHALLENGE · 03
Disconnected SIS / LMS / assessment
Roster in one system, grades in another, attendance in a third. Teachers spend 8h/week reconciling.
LTI 1.3 · OneRosterCHALLENGE · 04
Engagement signals lost in noise
Click-streams without context tell you nothing. Mastery velocity, hesitation, retries — those tell you everything.
Early-warning MLCHALLENGE · 05
Assessment integrity
Plagiarism, proctoring, AI-generated submissions. The integrity problem moved faster than most policies.
Liveness + driftCHALLENGE · 06
Student data protection
FERPA, COPPA, state student-data laws, GDPR for international cohorts. Every integration is a privacy decision.
PII minimizationWhat we build · and own
Six modules. One learning spine.
Bring your own LMS, SIS, content library or rubric — these modules slot in beside them with LTI 1.3 and OneRoster, and never ask you to switch.
Module · 01
Adaptive learning engine
Mastery-based pathways with item-response calibration. The next question is the one this learner needs, not the next one on the syllabus.
- Knowledge-graph backbone
- IRT difficulty calibration
- Mastery thresholds
- Spaced-repetition recall
Module · 02
AI tutor with curriculum guard-rails
LLM tutors grounded in your textbook, your worked examples, your rubric. They never invent prerequisites — and they hand back to a human at the first sign of confusion.
- Private retrieval (RAG)
- Citation-required answers
- Curriculum-aligned hints
- Teacher hand-off triggers
Module · 03
Engagement & early-warning analytics
Time-on-task, mastery velocity, hesitation patterns. Teacher dashboards that surface the at-risk learner this week, not after the term ends.
- Week-3 risk scoring
- Per-cohort heatmaps
- Parent-portal summaries
- Intervention playbooks
Module · 04
SIS · LMS · assessment glue
LTI 1.3, OneRoster, Common Cartridge, SCORM 2004, xAPI. We make the systems you have talk to each other without a manual import.
- Bi-directional sync
- SSO (SAML/OIDC)
- Roster + grade passback
- Universal gradebook
Module · 05
Proctoring & integrity
Liveness, gaze, room-scan, screen-share and AI-generated-content drift detection — applied only where the stakes warrant it.
- Drift detection for AI text
- Anomaly scoring
- Human-review queues
- Appeal workflow
Module · 06
Workforce skills & credentialing
For corporate L&D and continuing education — competency frameworks, skills graphs, verifiable credentials and ROI-grade dashboards.
- Skills graph · ESCO/OSI
- Open Badges 3.0
- LRS / xAPI store
- Manager dashboards
Reference architecture
From curriculum import to mastery telemetry, the path is six steps.
A single learner event — a hesitation, a wrong answer, a re-attempt — flows through six clean stops on its way to becoming a teacher's next action.
STEP · 01IngestCurriculum + rubric ingested. Vectorized. Linked to a knowledge graph.
STEP · 02AssessDiagnostic items calibrate baseline mastery per learner per concept.
STEP · 03AdaptPathway engine selects next item based on mastery + velocity.
STEP · 04TutorLLM tutor available on-demand, grounded in your content + rubric.
STEP · 05SignalHesitation, retries, time-on-task — surfaced to the teacher in week 3.
STEP · 06ReportMastery telemetry exported to SIS / parent portal / accreditation.
Tools we reach for
Battle-tested stack. No buzzword tax.
Application
Next.js · App RouterReact NativeTypeScripttRPCPostgrespgvector
AI / ML
Claude SonnetGPT-4oWhisperSentence-TransformersLangGraphEvidently AI
Standards & integration
LTI 1.3OneRoster 1.2xAPI / cmi5SCORM 2004Common CartridgeOpen Badges 3.0
Data & analytics
BigQuerySnowflakeLookerSegment · CDPIcebergdbt
Who we build for in education.
From early-years through doctoral programs, from corporate L&D to government workforce schemes — and the EdTech start-ups serving all of them.
K-12 districts & MATsHigher Ed · universitiesTest prep · admissionsEdTech start-upsCorporate L&DSkills training providersGovernment workforce schemesContinuing medical education
Case study · in production
How a K-12 group of 41 schools closed the engagement gap in one term.
Three thousand learners, eleven languages, one inherited LMS. We layered an adaptive math platform with an AI tutor that respected the state syllabus. By week 6, teachers had named at-risk learners they'd previously caught at semester end. By term end, course completion was up by a third.
ClientK-12 group · 41 schools
Learners3,142 active · grades 6–10
Timeline14 weeks to pilot
SubjectsMath · science · English
+34%Course completion · vs. prior cohort
12×Cheaper formative grading
−9 dAt-risk detection lead-time
92%Teacher adoption · 4-week
Compliance · built-in
Audit-ready by default. Not by sprint.
Student data is the most precious data we touch. We minimize, encrypt and segregate by default — and we don't train external models on your learners' work, period.
FERPA · USAStudent-record access + audit
COPPA · under-13Verifiable parental consent
GDPR · EU/EEALawful basis · data residency
SOC 2 · Type IIContinuous controls · evidence
State student-data lawsCA · NY · CO · TX exhibits
WCAG 2.2 AAAccessibility by default
LTI 1.3 · OneRosterIMS Global certified
Open Badges 3.0Verifiable credentials · W3C VC
Questions you’ll probably ask
The short version of the kick-off call.
Will an AI tutor replace our teachers?+
How do you keep the tutor from inventing things?+
Can you work with our existing LMS / SIS?+
How do you handle AI-generated student submissions?+
What does it cost to run?+
Do we own the data?+
Education · K-12 · higher ed · L&D
Bring your curriculum. We'll bring the engine.
A 60-minute working session with a senior engineer and a learning-science lead. We'll review your stack, your standards and your week-3 dropout problem.