Updates, tips, and stories.
Notes from the team building Cereby — what we're shipping, why, and how to get more out of your studying.
Hello Instead of an Answer: Auditing Cereby's Context Manager
A failed answer looked like a friendly greeting and nothing added up what the answering model read, so we audited Cereby's context manager and rebuilt it around one token budget, honest failures, and a prompt that finally sees the conversation.
The Gate That Never Fired: Auditing Cereby's Intent Classifier
Cereby's classifier is meant to ask when it is unsure what a student wants, and for a month it never asked once, so we audited it, found a confidence score built mostly from constants, and rebuilt it around the model's own judgment, in about half the prompt and inside a deadline.
Answered From the Preface: Auditing Cereby's File Compressor
Vague follow-ups about big files were answered from the preface; we found 22 flaws, and the answering passage now reaches the model 7 times in 8, not 3.
Pin the Assessment Again: Performance Analysis Without the Pin
Asking how you were doing in a subject showed a placeholder unless a quiz was pinned, so we had a model call find that subject's materials itself.
Page Unresponsive: Seeing Page Freezes in Our Error Dashboard
Chrome's Page Unresponsive dialog left no trace in our error dashboard, so we added a small heartbeat that reports each freeze there, grouped by route.
From 10MB to 500MB: Sending Cereby's Uploads Straight to the Bucket
Uploads over about 10MB died at the request limit before our code ran, so the browser sent files straight to storage and the cap rose to 500MB.
The Toggle That Felt Broken: Streaming Read-Aloud by the Sentence
Read-aloud waited for the whole reply before it spoke, so we streamed it a sentence at a time, and the first one played in roughly 500ms, not several seconds.
The Card Cereby Couldn't See: Enriching Pinned Study Materials
Pinned study materials reached Cereby without the per-card misses, tags and links we had stored, so we added a stage that sends them.
The Model Chose Not to Save It: Rebuilding Cereby's Memory
Cereby's chat model saved about 5 to 10% of the facts worth remembering, so a parallel classifier took over and saves rose to about 85 to 90%.
Seven Queries Nobody Read: Trimming Cereby's Orchestration Layer
File chat ran seven queries it threw away and recompressed files on every follow-up, so we classified first, cached the compression and prefetched embeddings.
A Test Became Flashcards: From First-Match-Wins to Parallel Scoring
A regex that answered first turned a request for a test into flashcards, so we made every classifier compete and let the best-scored candidate win.
Twelve Rules and a Paraphraser: Building the Cereby Humanizer
AI-assisted drafts read as machine-written, so we paired a paraphraser with 12 rules; on our sample the AI score fell from 68% to 51%, short of our 40% target.
46 MB for Three Pages: How We Cut Cereby's PDF Exports by 98%
A sparse three-page note downloaded as a PDF of over 46 MB; capturing less height, at a lower scale, as JPEG, brought it down to 1 MB.
The Proof on Page 12: From File Upload to Grounded Answer
Whole files in the prompt cost $1.50 to $2.00 a question on large PDFs, so we built a pipeline that sent the model only the part of a file a question needed.
Flaskcards, Not Flashcards: How We Fix Typos Without Wild Guesses
Cereby's fast-path matchers missed misspelled keywords, so we built a narrow normalizer that fixes routing words only when one answer wins.
A Port and a Key: Hardening Cereby's AI Detector API
Our AI detector ran on an open port with only a key in front of it, so we put Nginx, TLS, limits and bans around it, and an audit closed the port we missed.
Make This Punchier: Why Cereby Mini Moved to Two Intents
Cereby Mini's intent labels kept missing what students asked for, so we cut its first routing step to one question: change the page, or answer in chat.
62% and Nowhere to Look: Cereby's Standalone AI Text Detector
One score in the editor said a draft looked risky but not where, so we built a detector that scores each section and keeps every grade as a snapshot.
Stuck on Question 7: Introducing Cereby Tutor, AI Help Inside Quizzes
Asking another chat about a quiz meant retyping it and being handed the answer, so we built a tutor into each quiz that hints first and explains after.
One Line or Three Pages: Choosing an AI Model in Cereby
Every task ran on the same two models, mostly from one vendor, so we put a gateway in front of a multi-provider allowlist and let students choose a tier.
The Grab-Bag Prompt: Cereby's Four-Layer Context Model
Ad hoc prompts drifted in shape and were trimmed by guesswork, so we gave Cereby an ordered stack of context layers whose order is the trim policy.
Copy, Switch, Paste Back: Introducing Cereby Mini, Our Document AI
Pasting AI suggestions back into a document was slow and hid what changed, so we built an assistant inside the document that shows a diff first.
Citing Pages the Model Never Saw: Two-Stage Citation Verification
Compression shows the model 5 to 8% of a document, so a second pass now checks every quote against the full original and attaches the page it is really on.
Everything but the Answer: Query-Aware Compression for Oversized Pages
Cutting an oversized page at its token budget silently dropped the answer, so we now score the page's chunks against the question and send the best ones.
The $1.50 Question: Hierarchical Context Compression in File Chat
File Chat sent the whole document with every question, at $1.50 to $2.00 and 15 to 20 seconds, so we built compression that cut its tokens by 92%.
Five to Eight Seconds for a Follow-Up: Making Cereby AI Faster
Cereby AI took 5 to 8 seconds per request, so we cached its context, sent the model less and ran queries in parallel: 1 to 2 seconds on a cache hit.
The Question Nobody Typed: Dictation and Podcasts in Cereby
Typing on the move killed long questions, so we added push-to-talk dictation and generated podcasts: students speak to Cereby and listen back.
Cells Could Mean Three Subjects: Inside Cereby's Intent Classifier
A keyword router guessed which tool a student wanted, so we built a layered intent classifier that took accuracy from about 77% to about 95%.
Flat Bullets and Broken Math: One Visual Stack for Students and AI
Cereby AI had sound ideas but delivered flat bullets and broken math, so we gave it the same editor, math and templates students write with.
Three Failed Quizzes, Four Days Left: Introducing Cereby AI
A chat that forgot a student's quizzes and deadlines gave generic answers, so Cereby AI sent the model what the student had been doing, on every request.
The Chapter That Evaporated: Spaced Repetition and SM-2 in Cereby
A chapter read once is mostly gone within a week, so Cereby's flashcards use SM-2 to schedule each review just before you would forget it.