DeweyLearn Raises $5 Million to Enter AI Skills Assessment, Betting on Multimodal AI for Practical Skills in Cooking, Nursing, and Clinical Treatment
Education technology company DeweyLearn has completed a $5 million Series A funding round, betting on multimodal AI to assess practical skills in cooking, nursing, and clinical treatment, rather than just answering questions, lesson preparation, or grading multiple-choice tests. SJF Ventures led the round, with participation from Catalysis Capital, Morningside, and Owl Ventures; the funds will be used to expand into medical education, higher education, workforce training, and K-12 scenarios.
DeweyLearn was co-founded by Luyen Chou, former Chief Product Officer at Pearson and former Chief Learning Officer at 2U, along with AI expert Dirk Liebich. Its system combines video, audio, learning data, curriculum requirements, domain knowledge, and learning science to conduct preliminary analyses of learners' performance in real or simulated environments: not only determining whether tasks are completed but also attempting to identify action steps, operational sequences, communication methods, decision-making processes, and whether they meet the skill standards defined by educational institutions.
The company does not use a standardized universal scoring model but builds customized knowledge graphs and assessment frameworks for partner institutions. Schools input curriculum objectives, competency standards, scoring rubrics, industry norms, and teacher experiences into the system, which then compares learners' actual performance against the institution's standards; in cooking scenarios, the system can assist in analyzing knife skills, processes, and final product preparation, while in nursing and clinical training, it focuses on operational norms, communication quality, intervention steps, and situational judgment.
DeweyLearn's most representative case is the Auguste Escoffier School of Culinary Arts. The platform has conducted preliminary assessments on over 20,000 student assignments, with teachers retaining the right to review, modify, and make final judgments; the company claims this workflow has saved instructors hundreds of hours in grading time and reduced feedback cycles from weeks to hours, allowing teachers to focus on one-on-one tutoring and complex teaching issues.
The company is also expanding into non-traditional school scenarios: exploring assessment product collaborations with Riverside Insights, supporting therapist continuing education in the NeuroAffective Relational Model clinical training system, and planning to enter corporate skills certification, medical training, vocational education, and K-12 classrooms. DeweyLearn aims to establish a "Learning World Model," gradually forming an underlying understanding of human skill learning and performance processes through cross-domain, multi-curriculum, and multi-population learning behavior data.
DeweyLearn previously stood out from over 3,000 education technology companies, receiving recognition at the 2026 ASU+GSV Cup. This award does not equate to commercial scale but indicates that its approach of "AI observing real learning processes" has been distinctly recognized from the education AI track primarily focused on chatbots and automated content generation; ASU+GSV defines it as a multimodal AI system capable of assessing learning performance in both online and offline environments.
In market mechanisms, vocational schools, nursing schools, medical training institutions, universities, corporate training departments, and assessment agencies are DeweyLearn's buyers; the platform offers faster skill feedback, larger-scale process evaluation, and teacher time release. The bottleneck in traditional practical courses is not the lack of teaching content but the insufficient manpower for expert observation and scoring; if AI can take on the first round of recording, comparison, evidence organization, and preliminary evaluation, it can reallocate limited teacher resources to feedback interpretation, anomaly cases, ethical judgments, and personalized guidance. Beneficiaries include educational institutions with clear scoring standards, teachers, and learners; those under pressure are training systems that rely solely on manual observation and struggle to scale skill assessments.
Source: Public Information
ABAB AI Insight
The first phase of educational AI mainly revolves around "knowledge interaction": intelligent Q&A, personalized practice, assignment grading, content generation, and learning planning. Its basic inputs are text, questions, multiple-choice answers, or code, and the assessment object is what learners "know". DeweyLearn targets the more challenging second phase that is harder to digitize: whether students can correctly complete tasks in real environments, adhere to procedures, and communicate and make decisions in uncertain situations. This shifts educational AI from "cognitive assessment" to "behavioral assessment."
This change is directly related to the long-standing structural issues in vocational education. Cooking, nursing, clinical treatment, technical jobs, and corporate training all require expert observation of actual operations; however, expert assessment is costly, slow in feedback, has limited scoring consistency, and is difficult to scale for large numbers of learners. Traditional exams can use standardized question banks to scale but struggle to answer whether "students truly master operational skills." DeweyLearn attempts to combine video, audio, live recordings, and course scoring rubrics into auditable performance evidence, allowing AI to first complete repetitive observations while reserving high-risk judgments for teachers.
The key to its technical route is not using a stronger visual model but making each school's "implicit teaching standards" explicit. General models can recognize knives, pots, hospital beds, stethoscopes, and human actions, but may not know how a particular nursing school defines sterile procedures, how a specific cooking course evaluates knife skills, or how a particular treatment training assesses whether an intervention is appropriate. DeweyLearn builds customized knowledge graphs for institutions, essentially transforming teachers' scoring language, industry norms, and curriculum objectives into a model-callable judgment framework; this is also part of why it can form a paid barrier more easily than general video analysis tools.
In terms of capital pathways, the company first establishes data and assessment capabilities in vertical scenarios with clear standards, recordable processes, and heavy teacher scoring burdens, then expands into medical, higher education, workforce training, and K-12. The 20,000 assignments from the Auguste Escoffier culinary school are not just a customer case but also accumulate closed-loop data on "task-action-scoring-teacher revision-learning outcomes." Teacher approval, modification, and rejection of AI's preliminary evaluations will continuously correct the model's understanding of specific course standards; thus, the company is not merely selling a one-time model call but building an assessment data asset that can be enhanced with random use.
Similar logic has appeared in language testing, programming assessments, and online proctoring: Duolingo improves language ability judgments through extensive learning behavior data, CodeSignal and HackerRank assess technical abilities through code submissions and test cases, while organizations like Prometric rely on standardized exams and proctoring processes to maintain certification credibility. DeweyLearn aims to enter areas that these systems have not fully covered—skills certification that requires "seeing, hearing, and understanding the operational process." Its competition with traditional exam companies is not just about grading speed but who can define the evidence standards for practical abilities.
This represents technological substitution, but the object of substitution is not the teachers themselves but the time-consuming, repetitive observation and preliminary evaluation work that teachers do. AI can expand feedback coverage but cannot inherently assume final responsibility in high-risk scenarios: whether nursing operations are safe, whether clinical communication is appropriate, and whether students should receive certification all require human expert review, appeals mechanisms, data governance, and clear boundaries of responsibility. Future competition in educational AI will not only look at whether models can provide scores but whether they can offer explainable evidence under teacher review, maintain scoring consistency, and safely deploy across different school standards.
ABAB News · Cognitive Laws
Knowledge can answer questions; skills must be seen.
AI replaces grading time; teachers retain judgment responsibility.
General models recognize actions; industry standards determine value.