Research Papers
Sudar and the Adaptive Learning Layer (ALP) are described in academic work. Below is the primary paper, its key contributions, the full abstract, and a formatted citation.
Learning That Remembers You
An Open-Source AI-Native Learning Platform and Plugin Architecture for Longitudinal Learner Modelling at Scale
Abstract
Traditional learning management systems (LMSs) deliver static, one-size-fits-all content with no longitudinal learner model and no adaptive tutoring. Intelligent tutoring systems (ITS) that do adapt remain either narrow-domain research prototypes or products disconnected from the course-hosting infrastructure most organisations already use.
We present Sudar, a fully open-source, AI-native learning system released under the Apache 2.0 licence, making three contributions: (1) the Sudar reference platform (LAMP), a working implementation unifying authoring, delivery, and intelligence around a persistent Digital Learner Twin, adaptive sequencing, six multimodal delivery formats, an AI tutor with longitudinal cross-session memory, bounded agent orchestration (SudarAgents), and consent-gated generative personalisation; (2) the Adaptive Learning Layer (ALP), a novel plugin architecture enabling these capabilities to be deployed as independently installable services on top of existing LMSs without requiring platform replacement; and (3) an economic analysis demonstrating the full capability set can be delivered at a per-learner AI infrastructure cost below $0.02 per month, exceeding a 99% reduction relative to both incumbent commercial LMS licensing fees and proprietary AI provider stacks.
The reference implementation is open source (Apache 2.0), grounded in a broad learning-science evidence base, and designed as an extensible bedrock on which the community can build additional modalities, intelligence plugins, and LMS connectors.
docs/research/ in the repo for the current draftThree primary contributions
Each contribution addresses a distinct gap: the architectural gap, the integration gap, and the economic gap.
LAMP: Reference Platform
A fully open-source, working implementation of an AI-native learning system that unifies authoring, delivery, and intelligence around a persistent Digital Learner Twin, adaptive sequencing, six multimodal delivery formats (text + read-along TTS, animated video, audio podcast, mindmap, flashcards, SCORM), bounded agent orchestration (SudarAgents), and an AI tutor with longitudinal cross-session memory and consent-gated generative personalisation.
ALP: Adaptive Learning Layer
A novel plugin architecture enabling Sudar's capabilities (learner memory, adaptive tutoring, next-best-action recommendations, and modality choice) to be deployed as independently installable services on top of existing LMSs (Moodle, Canvas, Blackboard) without requiring platform replacement. Architecturally aligned with Moodle 4.5's AI subsystem. Potential reach: 500 million registered Moodle users across 233 countries.
Economic Analysis
Empirically observed infrastructure cost data demonstrating a >99% cost reduction relative to both incumbent commercial LMS licensing fees and proprietary AI provider stacks. Full AI-native personalised learning delivered at $0.021 per learner per month (less than the cost of a single SMS) using open-weight models and zero-cost TTS. All figures verified against Q1 2026 provider pricing.
Capability comparison
Table 1 from the paper. Sudar + ALP compared with representative systems.
Cost column shows AI infrastructure only at 1,000 learners per month (Q1 2026).
| Feature | Sudar + ALP | Khanmigo | LearnLM (Google) | Typical LMS + AI |
|---|---|---|---|---|
| Learner model | Longitudinal Digital Twin | Session-scoped | Stateless | None |
| Tutor memory | Cross-session, cross-course | Within session | Not claimed | Stateless |
| Open learner model | ✓ (inspectable + editable) | No | No | No |
| Modalities | 6 (text/listen/video/podcast/map/cards) | Text | Text / slides / audio / map | Text / video |
| Augments existing LMS | ✓ (ALP) | No | No | N/A |
| Open source | ✓ Apache 2.0 | No | No | Rarely |
| Cost (1,000 learners/mo) | ~$21 | ~$4,000 | N/A | $3,400-$44,000 |
Infrastructure cost at 1,000 learners / month
Table 3 from the paper. Empirically observed Q1 2026 pricing. AI infrastructure only; hosting costs excluded.
Supabase free tier: $0. Studio / Learn on Vercel free tier: $0. Intelligence on Railway / Render: about $5-$10/month at moderate traffic.
| Stack | Per learner / mo | Annual (1,000) |
|---|---|---|
| Sudar (Together AI 8B + Edge-TTS) | $21 | $252 |
| Sudar (Together AI 70B + Edge-TTS) | $105 | $1,260 |
| Self-hosted (Ollama + open-weight, GPU) | ~$0* | ~$0* |
| GPT-4o + OpenAI TTS-1 | $3,405 | $40,860 |
| Claude 3.5 Sonnet + Azure TTS | $3,735 | $44,820 |
| Docebo (platform licence, low end) | $5,830+ | $69,960+ |
| Sana Labs (est., min. 300 seats) | ~$15,000 | ~$180,000 |
*Self-hosted ~$0 per API call after hardware provisioning; excludes amortised GPU hardware. Sudar achieves a >99% cost reduction relative to both proprietary AI stacks and incumbent LMS platform fees.
Cite this paper
When using Sudar or ALP in research or derivative work, please cite:
BibTeX
@misc{karunanithi2026sudar,
author = {Karunanithi, Dhanikesh},
title = {Learning That Remembers You: An Open-Source
{AI}-Native Learning Platform and Plugin Architecture
for Longitudinal Learner Modelling at Scale},
year = {2026},
url = {https://github.com/Dhanikesh-Karunanithi/Sudar},
note = {Sudar / ALP Project. Apache-2.0 licence.
arXiv preprint pending.}
}APA
Karunanithi, D. (2026). Learning That Remembers You: An Open-Source AI-Native Learning Platform and Plugin Architecture for Longitudinal Learner Modelling at Scale. Sudar / ALP Project. Apache-2.0 licence. arXiv preprint pending. https://github.com/Dhanikesh-Karunanithi/Sudar
Reproducibility artefacts
All documents referenced in Appendix C of the paper are available in the repository:
docs/ALP_API.mdALP HTTP endpoint reference, field mappings, and integration examplesdocs/AGENTS_PLATFORM.mdSudarAgents mission catalogue, tool catalogue, and security modeldocs/research/COST_WORKSHEET.mdVersioned cost assumptions; fill and update at time of deploymentdocs/research/EVALUATION_APPENDIX.mdScope and limitations for each empirical claimdocs/research/PILOT_PROTOCOL.mdIRB-ready pilot study protocolscripts/benchmark-sudar.mjsAutomated latency and token measurement scriptECOSYSTEM.mdFull schema, event model, and API surfaceRESEARCH_FOUNDATION.mdEvidence-to-feature mapping (37 citations)