Trilogy · Alpha School · StudyReel
A Trilogy portfolio build · 8 weeks to productionThe AI tutor that watches, understands, and coaches — built on Claude vision
Trilogy, the PE firm behind Alpha, wanted to retire the last human bottleneck in their 2 Hour Learning model: the in-room guide. StudyReel is what replaced it. Three live streams — screen, camera, audio — fuse into one understanding of every student’s learning state, powered by Claude’s vision models running on a tight latency budget.
The Trilogy Thesis
Trilogy runs a portfolio built around one conviction: every category has a version where AI does the job a human used to. Alpha’s schools had already proven the academic outcomes — top 2% nationally, 2.3× the national average pace. What hadn’t scaled was the human attention layer. StudyReel is the AI that sits in that seat, simultaneously, for every student.
Alpha School had already proven that students could master a full day's academics in just two hours using adaptive software. But the model depended on in-room guides to keep students focused and on track. To scale beyond a handful of campuses, they needed something that could do what a great tutor does — observe, understand, and intervene in real time — without a human in the loop.
What We Built
An AI that actually sees the student
StudyReel captures three streams simultaneously: the student's screen, their camera feed, and audio. Vision models analyze what's happening in real time — is the student engaged with the material? Stuck on a problem? Distracted? Off-task entirely?
This isn't surveillance. It's the digital equivalent of a tutor sitting next to you, watching your work, and knowing exactly when to step in. The system understands context — the difference between productive struggle and genuine confusion — and responds accordingly.
Real-time AI coaching
When the system detects a student struggling, the AI Coach intervenes with personalized guidance calibrated to that student's level, learning style, and current context. It doesn't give answers. It asks the right questions, offers hints, and nudges students toward mastery.
The coaching adapts continuously. A student who responds well to visual explanations gets diagrams. One who needs to talk through problems gets a Socratic dialogue. The AI builds a model of each learner and refines its approach with every session.
Two hours to outperform eight
The traditional school day is eight hours long. Alpha students finish their academics in two. Not because the curriculum is lighter — because the learning is denser. Every minute is personalized. No waiting for 25 other kids to catch up. No re-teaching material already mastered.
MAP Growth assessments confirm the results: Alpha students progress at 2.3× the national average, with the top 20% achieving 6.5× growth. The remaining six hours? Spent on real-world skills, sports, arts, and projects that traditional schools squeeze out of the schedule.
The Claude Stack
Vision Tier
Sonnet · multimodal
Frame-sampled screen + webcam fused into a single prompt every few seconds. Picks up confusion, distraction, and productive struggle from posture and pacing, not keystrokes.
Fast Tier
Haiku · triage + gates
Every stream tick starts on Haiku. It decides whether the session looks normal, needs a gentle nudge, or needs the coach. Most ticks stop here.
Coach Tier
Opus · Socratic generation
When the student is genuinely stuck, Opus generates the intervention — a question, a hint, a worked example — calibrated to this student’s learner model.
Orchestrator
Agent SDK · per-student sessions
Every student is a long-running session. Subagents handle vision, audio, and curriculum state; the supervising agent decides when to escalate.
Memory
Per-student learner model
The adaptive picture of each student — what explanations land, what tends to confuse, current mastery — persists across sessions and nudges Opus’s coaching style.
Caching
Curriculum prompt cache
The lesson context and rubric for each unit is cached, so a 45-minute session costs a fraction of what the naive approach would.
The Tutor Loop
From three raw streams to a single intervention, in seconds
Capture
Screen, camera, and audio stream into a per-student session. Frames and audio chunks are sampled, deduplicated, and batched into multimodal prompts.
Triage
Haiku reads the fused frame + transcript and emits a terse structured state: focused, confused, distracted, stuck, off-task. Most ticks return ‘focused’ and the loop rests.
Context assembly
On escalation, a subagent pulls the current lesson, the student’s learner model, and recent transcript. These are joined against a cached curriculum prompt.
Coach
Opus generates the intervention: a Socratic question, a targeted hint, or a worked example — in the modality (diagram vs dialogue) the learner model says this student responds to best.
Learn
The student’s response — did the hint land, did they get unstuck — is folded back into the learner model so the next intervention is better calibrated.
Why Claude
Safety-first vision, for a product pointed at children
StudyReel sits in front of K-12 students for hours a day. Trilogy and Alpha needed a model with vision capabilities they could trust, refusal behavior they could reason about, and a support story for the inevitable policy questions from parents and regulators. Claude was the model where those three were already in the box.
The ladder of Haiku → Sonnet → Opus is also what makes the economics work. If every tick ran on the coaching model, the margin wouldn’t exist. Triage-first is the shape that lets a network of schools run on a per-seat budget.
The Shift
From guides in a room to intelligence at scale
Alpha's original model replaced teachers with guides — adults who kept students focused while software handled instruction. It worked, but it didn't scale. Every campus needed guides. Every guide needed training. The bottleneck wasn't the curriculum. It was the human attention layer.
StudyReel removes that bottleneck. The AI observes what guides used to observe, coaches where guides used to coach, and does it for every student simultaneously. Not a replacement for human connection — a replacement for the repetitive monitoring that kept the model from scaling to thousands of students across multiple states.
Under the Hood
Three streams, one understanding
Screen capture shows what the student is working on — which problem, which app, how long they've been stuck. Camera feed reads engagement signals — focus, frustration, distraction. Audio captures questions asked aloud and ambient context. The AI fuses all three into a real-time picture of each student's learning state.
This multimodal approach is what separates StudyReel from simple adaptive software. Traditional edtech knows if you got the answer right. StudyReel knows if you guessed, if you're about to give up, and whether you actually understood the concept or just pattern-matched the solution.
Scale
A new category of school
Alpha started as a single campus in Austin. The 2 Hour Learning model now powers a growing network of schools — Alpha, Unbound Academy, and partner campuses across multiple states. Each school runs the same core platform but adapts to its student population.
The Wall Street Journal, FOX News, and Forbes have all covered the model. Not as a curiosity — as a signal that the fundamental structure of schooling is changing. When students consistently outperform their peers with a fraction of the seat time, the question stops being “does this work?” and becomes “why are we still doing it the old way?”
Top 2% nationally. 2.3× faster progress. Two hours a day. This is what happens when the model carries the classroom.
Built on Claude. Designed to compound.
LightCI partners with companies shipping production agents on the Claude Agent SDK.
Talk to LightCI