Talimo
A full AI-powered medical education platform — designed, built, and shipped solo in 4 weeks. From lecture upload to clinical simulation, one platform that replaces five separate tools.
0 weeks
Time to ship
0+
Active students
0+
Flashcards created
0+
Lectures uploaded
The Problem
Med students spend more time preparing to study than actually studying.
My partner is a medical student. I watched her spend 3 hours making flashcards from a 1-hour lecture. Then she'd buy expensive question banks that ran out of questions. Then she'd go to clerkships in year 3 having never practiced taking a patient history with real feedback.
The tools existed — Anki for flashcards, UWorld for questions, YouTube for lectures, Google Docs for notes — but nothing talked to each other. There was no unified view of what you were weak on. The students who thrived weren't necessarily the ones who understood medicine best — they were the ones who were good at organizing their studying.
A typical Tuesday night
8 hours. One lecture. Nothing connected. Nothing retained.
Same Tuesday with Talimo
The Insight
Automate the organization layer so the actual learning becomes the bottleneck — not the prep.
Every med student has the same 24 hours. The difference between students who keep up and students who fall behind isn't intelligence — it's how much time they lose to busywork. If I could make flashcard creation, question generation, and clinical practice happen automatically, students could spend all their time on the thing that actually matters: learning medicine.
What I Built
An AI study companion that learns how you learn.
Most study tools are silos. Your flashcard app doesn't know you just bombed a cardiology quiz. Your question bank doesn't know you keep missing the same concept in case simulations.
Talimo is different. Meet Neu— a friendly AI tutor that lives across every feature. It remembers what you've studied, what you're struggling with, and what you're about to forget, then adapts your entire study experience around that.
Everything starts from a lecture.
Meet Neu
Upload a lecture. Get everything.
Drop a PDF, PPTX, audio file, video, or YouTube link. In under a minute, Talimo generates a study guide organized by learning objectives, a flashcard deck, 3 tiers of practice quizzes, cascade pathways, and a concept map. Embedded images are extracted, analyzed by Claude Vision, and matched to the relevant sections.
Then Neu is available to chat — with full context of the lecture content andyour performance history on related topics. It's not a generic chatbot. It knows what you got wrong last week.
Heart_Failure_2026_McCully.pdf
2.7 MB — Ready to upload
Flashcards that know when you're about to forget
Three card types — basic, cloze deletion, and image occlusion where Google Cloud Vision detects labels in anatomy diagrams and generates a card per structure. An SM-2 spaced repetition engine schedules reviews at the optimal moment.
When you get a card wrong, Neu doesn't just show the answer. It explains why, generates a memory hook with a visual recall chain, and surfaces related cards you should review next. Full Anki import up to 2GB and community deck sharing.
Question
What nerve innervates the diaphragm?
Answer
The phrenic nerve (C3, C4, C5).
Unlimited questions, weighted to your weak spots
Type any topic and get board-style clinical vignettes with per-option explanations. Auto-detects your exam format — USMLE, COMLEX, NCLEX, PANCE, or NAPLEX — from your profile. Six question types from basic recall to two-step reasoning.
Neu uses your full performance history — flashcard retention, case sim grades, past quiz scores — to weight questions toward the areas where you're actually weak. You're never wasting time on topics you already know.
Your question:
Which medication class reduces mortality in HFrEF by blocking the RAAS pathway at the ACE level?
Practice on AI patients before you see real ones
A 5-phase clinical encounter: interview the patient, build a differential, defend it under Socratic questioning from an AI attending, commit to a diagnosis, and write a SOAP note under a 10-minute timer. Voice mode lets students practice history-taking out loud with text-to-speech responses.
The system grades across 5 AAMC-mapped competencies — history gathering, clinical reasoning, patient communication, professional presentation, and self-directed learning — with cited evidence pulled directly from the conversation. Before clerkships, this is the closest thing to ward rounds.
Marcus Thompson, 34M
Chief Complaint: Chest Pain
What makes it stick
Everything feeds into a unified dashboard — cards due, study streaks, upcoming deadlines, a GitHub-style activity heatmap across all features, and per-lecture status tracking computed from flashcard mastery, quiz scores, and pathway completion. A gamification layer with 37 badges across 8 categories gives students milestones to chase.
And because medical education shouldn't be isolating, there are 3 real-time multiplayer game modes — quiz races with streak multipliers, diagnostic reasoning chains with peer voting, and clinical pathway sequencing — up to 6 players per room. It turned out to be one of the most-loved features: study groups use it as a pre-exam ritual.
How I Built It
Solo. Four weeks. AI-assisted throughout.
I built Talimo entirely by myself, using AI coding tools to move at a pace that would normally require a team. I played every role — product manager, designer, engineer, and QA — making hundreds of prioritization decisions daily.
The approach was simple: de-risk first, then compound. I started with the hardest technical problem (streaming AI conversations that feel real), proved it worked, then layered on features that built on that foundation — each one feeding data into Neu's cross-feature context.
Week 1
Foundation
- Database schema design and auth (Supabase + Prisma)
- Core AI interaction layer — streaming SSE pipeline with Claude
- Railway worker service for CPU-heavy media processing
Week 2
Core features
- Lecture upload pipeline — PDF, PPTX, audio, video, YouTube ingestion
- AI content generation — study guides, flashcards, quizzes, concept maps
- Spaced repetition engine (SM-2) with 3 card types
- AI case simulations — 5-phase encounters with AAMC competency grading
Week 3
Depth & intelligence
- Neu AI tutor — cross-feature context, per-lecture chat, explanations
- Question bank with exam-style detection and performance-weighted generation
- Anki import (chunked upload, up to 2GB) and community deck sharing
- Image occlusion cards via Google Cloud Vision
Week 4
Engagement & polish
- Real-time multiplayer games (3 modes, WebSocket-based)
- Dashboard with activity heatmap, streak tracking, and badge system
- Freemium pricing with Stripe integration
- Marketing site, analytics, and launch
Key Decisions
The tradeoffs that shaped the product.
Proved the lecture pipeline first
If the AI couldn't turn a raw lecture into genuinely useful study material, nothing else mattered. I focused on getting that right before building anything downstream — flashcards, quizzes, case sims all depend on the quality of that initial content generation. Getting it to produce meaningful, structured output from messy PDFs and hour-long recordings shaped every architectural decision that followed.
Separate worker service on Railway for media processing
Processing lecture videos with yt-dlp and ffmpeg is CPU-heavy and slow. I could've done it in Next.js API routes, but it would've blocked the event loop and made the entire app feel sluggish. Offloading to a dedicated Express service on Railway kept the main app responsive and let me scale processing independently.
Claude for AI generation, Whisper for transcription
I used each model where it's strongest. Claude handles all the educational content generation — patient dialogues, explanations, question writing — because it's exceptionally good at nuanced, structured medical reasoning. Whisper handles audio transcription from lecture videos because it's the best transcription model available. One model per job, picked for the job.
Supabase for auth + database instead of building custom
Four weeks means zero time for yak-shaving. Supabase gave me Postgres, auth, row-level security, and real-time subscriptions out of the box. I spent my time on product logic instead of infrastructure. The right amount of build-vs-buy for a solo sprint.
Studying shouldn't be a solo experience
Medical education is isolating — every existing tool is a solo experience. I deliberately built social into every layer: real-time multiplayer games for study groups, public deck sharing so students can follow people who create great material, user profiles with badges and streaks, and a community case library with ratings. The bet was that studying becomes sticky when it stops being lonely. It worked — study groups now use the multiplayer trivia as a pre-exam ritual.
Architecture
System design
A Next.js app backed by Supabase for data and auth, Claude for all AI generation, and a separate Railway worker for CPU-heavy media processing. Keeping the worker separate was key — lecture videos with yt-dlp and ffmpeg would have blocked the main app.
Claude API
Patient sims · Content gen
Student
Browser / Mobile
Next.js 14
App Router · React 18 · TypeScript
Supabase
PostgreSQL · Auth · RLS · Real-time
Railway Worker
Express · yt-dlp · ffmpeg
Whisper API
Lecture transcription
Student
Browser / Mobile
Next.js 14
App Router · React 18 · TypeScript
Claude API
Patient sims · Content gen
Supabase
PostgreSQL · Auth · RLS
Railway Worker
Express · yt-dlp · ffmpeg
Whisper API
Lecture transcription
Tech stack
Frontend
Backend & Infrastructure
AI & ML
Analytics & UX
What Students Say
From the people using it every day.
“I uploaded a 94-slide cardiology lecture and had a complete study guide, flashcard deck, and practice quiz in 30 seconds. That used to take me an entire evening.”
Sarah M.
MS2, Internal Medicine
“The AI case simulations are the closest thing to ward rounds I've found. I do voice-chat history-taking every night before bed. My clinical reasoning has gotten noticeably sharper.”
James K.
MS3, Clinical Rotations
“Our study group uses the multiplayer trivia as a pre-exam ritual now. It's the only time studying actually feels fun.”
Priya R.
MS1, Foundations
What I Learned
Lessons from building a full product alone in 4 weeks.
Building Talimo taught me that the hardest part of product management isn't coordination — it's judgment. When you're the only person making decisions, there's no one to blame and no committee to hide behind. Every feature you build is a feature you're not building. Every hour spent on infrastructure is an hour not spent on UX. You feel every tradeoff in real-time.
The biggest surprise was how much the “boring” decisions mattered. Choosing Supabase over a custom auth system saved me a week. Building the Railway worker early meant I never hit a performance wall. The architecture decisions I made in Week 1 determined what was even possible by Week 4.
But the real lesson is about AI products specifically: the model is maybe 20% of the work. The other 80% is the interaction design — making a non-deterministic system feel trustworthy, useful, and worth coming back to. Students don't care that it's Claude under the hood. They care that the AI patient feels real and the flashcard explanations actually help them learn.
Want to see it live?
Talimo is live and serving 500+ medical students. Try the platform or read more about how it was built.