A photograph of a plated dinner lands in an instructor's queue. Somewhere off-camera, an online culinary student has diced, sautéed and arranged the evidence of a lesson. The usual bottleneck is human and familiar: a qualified instructor has to inspect the work, compare it with a rubric, write useful comments and move to the next plate. DeweyLearn wants to take the first shift.
At Auguste Escoffier School of Culinary Arts, the New York startup says its system has reviewed more than 20,000 homework submissions, preparing assessments for instructors to approve or adjust. The point is not robot Gordon Ramsay. It is the less television-friendly business of returning hundreds of grading hours to faculty while giving students feedback sooner. That is DeweyLearn's most persuasive demonstration because it makes the product legible in one sentence: the machine studies the work, the expert keeps the final say.
The scarce product is judgment
Most education software begins with content: lessons, quizzes, flash cards, answers. DeweyLearn begins somewhere harder. It asks an institution to define what good performance looks like, including the standards filed in documents and the instincts sitting inside experienced practitioners. The platform structures those rubrics, examples and judgments into a custom knowledge graph. It then evaluates a recorded performance, live session or submitted artifact against that local definition of quality.
That sequence matters. A generic language model can sound knowledgeable while missing the standard that actually governs a nursing simulation, a therapy intervention or a consulting presentation. DeweyLearn's proposed difference is specificity: the customer's experts build the frame; multimodal AI reads the video, audio, behavior and work product; feedback points back to observed evidence. Over time, recurring strengths and gaps become visible across a learner or cohort.
The product in four moves
The commercial model follows the technical one. This is B2B enterprise software, configured for an institution rather than downloaded by an individual learner. Public pricing is not listed. The real cost is therefore not only a license; it includes the experts' time to make tacit standards explicit, the integration work, governance and ongoing review. A clean rubric and a library of good examples make the machine useful. A disputed rubric simply lets software reproduce an argument faster.
“We're making the kind of expert feedback that once required a master watching over your shoulder available to every learner.”Luyen Chou, co-founder and CEO
The pivot hiding in plain sight
DeweyLearn launched in early 2025 with a classroom-centered story: combine audio, video and learning data to reveal student comprehension, cognitive demand, engagement and collaboration. Its case study with Meteor Education applied custom observation frameworks to teacher coaching and prosocial behavior. The problem was real, but the pitch risked drifting into the crowded and sensitive territory of classroom analytics, where every engagement score invites a question about bias, consent and what the camera failed to understand.
By 2026, the company had widened the aperture. The current language is about performance assessment across workforce, clinical and corporate learning. Culinary photographs, therapist training and simulated clinical work all fit the same engine. This is not evidence that the first idea failed, and the company has not described it that way. It is evidence that the founders found a larger and more concrete job for the machinery: not merely describing a classroom, but helping an expert evaluate work that already demands review.
Luyen Chou brings unusually deep edtech scar tissue to that choice. Before DeweyLearn, he was chief learning officer at 2U, chief product officer at Pearson and a co-founder of both Prospect Schools and The School at Columbia University. Co-founder Dirk Liebich, president and CTO, supplies the applied AI, predictive systems and knowledge-graph architecture. The company itself is named for John Dewey, the philosopher who argued that people learn by doing. There is a family footnote worthy of a novel: during Dewey's travels in China a century ago, he encouraged Chou's grandmother to study teaching at the University of Chicago.
A wedge, then a world model
The founders' ambition is bigger than faster grading. Liebich has compared the destination to a kind of Google Earth for learning: customer-specific graphs accumulate cycles of observation, intervention and measurement, while a broader model discovers patterns in how people learn. That is a compelling research direction and a demanding business promise. Enterprise buyers usually purchase a painful workflow before they purchase a world model.
Where DeweyLearn sits
Text-first answers
Observed performance + local standards
Rich judgment, scarce time
Clicks, scores, completion
Review and reflection
Standardized tasks
That position puts DeweyLearn beside several kinds of alternative rather than one neat competitor. An institution can keep paying instructors to observe and grade manually. It can use learning-management dashboards, video-coaching products such as Edthena or Sibme, skills-assessment tools such as Bongo, or broad AI education platforms. DeweyLearn's edge is the combination: multimodal evidence, a custom institutional framework and feedback that can travel from one learner to cohort-level patterns.
The early customer list is correspondingly eclectic. Escoffier supplies the most mature public example. Meteor Education tested classroom observation. The NeuroAffective Relational Model is using the platform in continuing education for therapists working with developmental and complex trauma. Riverside Insights has said it is exploring new assessment approaches with DeweyLearn. Teachers College reported that more than 2,000 culinary students were receiving targeted feedback, while consultants and clinical trainees were using the technology in other settings.
Trust is the actual go-to-market
In April 2026, DeweyLearn won Top Startup and Best Workforce at the ASU+GSV Cup, selected from more than 3,000 nominees. Three months later it announced an oversubscribed $5 million Series A led by SJF Ventures, with Catalysis Capital, Morningside and Owl Ventures participating. The round gives a roughly six-person company room to improve the assessment technology and enter more domains. It does not settle the central question.
Multimodal learning data is intimate. A system that sees posture, hears hesitation and infers behavior can feel like a patient coach or an invisible manager. The difference lives in consent, access, retention, explainability and power. DeweyLearn says customer knowledge remains private, video capture can run on local devices and private cloud, outputs supplement human judgment, and assessments are grounded in observed behavior. Those are useful design commitments. Buyers will still need to test error patterns across populations, give people a meaningful way to challenge an assessment and decide where recording should never happen.
The product is least likely to work when performance cannot be captured cleanly, standards are ambiguous, rare context determines the right judgment, or users reasonably distrust the observer. It is also a poor fit when the review volume is too low to justify customization. The sweet spot is almost the reverse: repeated, observable work; a stable rubric; expensive expert attention; delayed feedback; and an institution willing to invest in governance.
There is another constraint hiding inside the word “expert.” Institutions must decide whose judgment becomes the system's judgment. Senior practitioners often disagree, and yesterday's standard can preserve yesterday's bias. The healthiest deployment would treat rubric-building as product work: compare assessors, document exceptions, sample the machine's misses and let learners see the basis for consequential feedback. Under those conditions, automation can make a standard more inspectable. Without them, a knowledge graph may merely give institutional habit a technical costume.
“The best applications of AI are the ones that continue to put humans in the loop.”Luyen Chou
That line is easy to print and hard to operationalize. DeweyLearn's opportunity comes from making it literal: instructors approve or adjust the culinary review; an organization's experts define the evaluation frame; evidence travels with the recommendation. If the company can preserve those mechanics as volume and domain count rise, it may have found a durable role for AI in education - the tireless first observer, not the final authority.
The useful lesson for another founder is smaller than the world model and sharper than the award. Find a queue where expert judgment is both precious and repetitive. Turn the expert's tacit standard into an explicit system. Give the machine the boring first pass. Give the human the consequential last word. Then measure time saved, feedback speed and whether learners actually improve. DeweyLearn has evidence for the first two. The next chapter is proving the third.