Waking up our servers

They were taking a quick nap

Orca21 Labs

Towards RSI for the $1.8T U.S. learning economy

Orca21 Labs is building self-improving learning systems that make expert, one-on-one teaching as cheap as software.

An abstract diagram of our system. The learner's work flows into their harness and into their profile. The strategy model reads the profile and proposes a change to the harness. A reviewer checks the change, and the harness then changes shape, always toward the goal.

  • 150+ points

    Early results have shown an improvement of about 150+ points in SAT scores for our first students in Learn 0.1a.

  • Google and MIT

    Our CTO is first author on AI research with both.

The Orca21 Labs memoRead the full memo →

Something is structurally wrong with how AI is used in learning

A student preparing for the SAT, a nurse learning a new protocol and an engineer moving to a new programming language each learn best with a plan built around their knowledge, goals and deadlines. A good teacher builds that plan, but that kind of teaching is rare and expensive. Learning has changed in stages, and each stage made learning fit the person a little better. The diagram below steps through the stages in order, and the red line under each stage says what that stage still misses. None of the first three stages builds a plan for one person and learns from results which plans work.

Learner ALearner BLearner COne coursesame lessons, same orderthe same path for everyoneA paid expert$$$plans around one learner, by the hour

Stage 1 · Textbooks, classes, online courses

One course for everyone

Every learner takes the same course, in the same order. A few families also pay a human expert to plan around their child.

Missing: A plan made for one person is rare and costly.

Our goal: recursive self-improvement for learning

We want an AI system that teaches better with every learner it works with, and that also gets better at the way it improves.

Human experts define what good teaching looks like. The system will learn from their corrections, and over time it will need them less.

It works as three loops, one inside the other.

Three loops, one inside the other, around one learner. Loop 1: after every session, the strategy model proposes changes to the learner's harness. Loop 2: Pod, which we are building, will compare results across all learners and give the changes that help to every harness. Loop 3: the strategy model proposes changes to the way it chooses changes, and a reviewer approves each one. Under the loops, a line of check marks: today a reviewer checks every change, and over time fewer changes will need a check.

We want it to serve the whole learning economy, one step at a time.

Four steps on one shared engine, each larger than the last. Tutoring is live today. School all year, anything you want to learn, and work (job training, onboarding and upskilling) are being built. Arrows show what each step will pass on to the next.

Every step will run on the same engine, so what it learns in one area carries over to the others.

We are starting with tutoring.

Read the full problem and our goal in the memo

Where we are right now

The road to RSI?Recursive self-improvement. A system that gets better at its job, and also gets better at the way it improves itself. is long, so we are building toward it step by step. The first step is the Learn 0.1 suite: two versions of our system that run inside LessonOrca, our consumer brand. Both use the same harness?Everything one learner gets: their AI agents (such as a tutor and a quiz maker), their plan and their lesson formats. and strategy model?A second AI for each learner. After every session, it reads the learner's results and proposes changes to their harness., and in both a reviewer approves, corrects or rejects every change. They differ in how much reviewer time each learner gets. Learn 0.1a shows how good results can be when a reviewer checks every change, and Learn 0.1b will test how close we can get with far less reviewer time.

  • Live

    Learn 0.1a

    Our system inside LessonOrca tutoring, where students also meet a human tutor every week. A reviewer approves, corrects or rejects every change the strategy model proposes.

    Early results have shown an improvement of about 150+ points in SAT scores for our first students.

    See LessonOrca tutoring Read about Learn 0.1a in the memo
  • Launching soon

    Learn 0.1b

    A LessonOrca app for the SAT, PSAT, ACT, AP and subject tutoring, where an AI tutor teaches every lesson. A reviewer checks each learner’s plan from a control center.

    One reviewer will look after 75 to 100 learners, at about 10 minutes per learner each week.

    See the Learn 0.1b page Read about Learn 0.1b in the memo

Coming soon

After the Learn 0.1 suite, we will build Pod 1 and then Learn 1. Each one starts once the step before it meets its target.

  • Coming soon

    Pod 1

    The part of our system that will learn from all learners at once. With learners’ consent, Pod will compare results across learners, find the changes that helped, and add them to the starting harness that every new learner gets.

    Read about Pod in the memo
  • Coming soon

    Learn 1

    Our goal: the first recursively self-improving harness for learning anything. A learner will enter their goals and constraints, such as a deadline, and Learn 1 will set up their AI agents, activities and progress tracking, then keep improving them for that learner.

    See the roadmap in the memo
Read where we are and our roadmap in the memo

Team

Abhiraam Eranti

Co-founder and CTO

Self-supervised learning, foundation models and learner modeling

  • First-author paper with Google and MIT (IEEE Access, 2026): a self-supervised foundation model for spectroscopy that trains with 60–90% less labeled data

  • Technical lead, U.S. Treasury EyeNote: built its currency recognition model (99.9% precision) for 100,000+ blind and low-vision Americans, starting in high school

  • Inventor on a provisional patent in impedance spectroscopy: built Chronus Health’s best 3-part white blood cell classifier, in use for over a year

  • Architect of the LessonOrca Student Profile: turns hundreds of observations about a student into one context record

  • Built ZemoMemo: models each learner’s forgetting curve (FSRS-6) and schedules recall for 1,000+ learners

  • UC Berkeley, Data Science
  • Berkeley SkyDeck, Entrepreneur in Residence
  • 1570 SAT
  • National Merit Finalist
GoogleMITU.S. Bureau of Engraving and PrintingUC Berkeley

Waleed Chaudhry

Co-founder and CEO

AI infrastructure, education businesses and go-to-market

  • AI/ML consultant at Databricks: infrastructure research with field engineering and product: lakehouse architectures, open table format benchmarks, enterprise AI strategy

  • Founder and CEO of WEInstruct, an SAT and ACT tutoring company: personalized plans for 30+ students, 5 tutors, $10,000+ revenue

  • Founder of Tin2Totes, a profitable recycling operation: 50+ collection sites across Washington and Oregon, recurring monthly revenue

  • Founder of The Bag Project: raised $10,000+ and distributed 10,000+ units across four states and Pakistan

  • Washington DECA State Vice President: elected to lead 15,000+ members across 15 regions; International Finalist and State Champion, National Economics Challenge

  • UC Berkeley, Haas School of Business
  • Coca-Cola Scholars Semifinalist
  • 1540 SAT
  • 35 ACT
DatabricksDECAUC Berkeley Haas

Talk to us

Email us at [email protected].

Read the full memo See the LessonOrca page
Orca21 Labs Orca21 Labs is the team behind LessonOrca.

SAT® and AP® are trademarks registered by the College Board. PSAT/NMSQT® is a registered trademark of the College Board and the National Merit Scholarship Corporation, which are not affiliated with, and do not endorse, LessonOrca. PSAT™ 10 and PSAT™ 8/9 are trademarks owned by the College Board. ACT® is a registered trademark of ACT Education Corp. None of these organizations is affiliated with, or endorses, LessonOrca or this site. Other product names belong to their owners.