Blog
Exam news, revision advice, and what changed in the latest release.
VTU's grace marks policy is real, dated 25 May 2017, and almost nothing students repeat about it is accurate. Here is what the notification actually says.
VTU declared its sixth-semester results five minutes after the last practical exam in June 2026. Here is how the turnaround works, where to check your result, and exactly what revaluation costs if a subject goes badly.
In VTU's 2025 scheme first year is nearly the same for every branch. What changes is your cycle. Here is the full subject list, with codes.
VTU needs 75% attendance in each subject. The useful question is not what the rule is, but how many more you can miss — and that has an exact answer.
Codes like 1BMATC101 and 1BESC104A are not random. Here is what each part of a first-year 2025 paper code tells you, read off the catalogue itself.
Three internals are conducted and two of them count. Here is how the best two are picked, how the average is taken, and why the third is not wasted.
VTU publishes one conversion and it is not CGPA × 10. Here is the formula, where the 0.75 comes from, and what it does to a placement cut-off.
Daily Revision schedules reviews on a five-rung ladder and then labels them with times that cannot be right. An item untouched for a month still reads "Yesterday".
Flashcards has three states and no spaced-repetition scheduler. Marking a card known removes it from revision permanently, and four different screens disagree about which cards are due.
Daily Quiz seeds nothing by date and generates nothing when the app opens. The quizzes are the output of a durable background queue, and the only daily thing about the feature is which five of them we remind you about.
AI Quiz lets you choose 10, 15, 20, 25 or 50 questions. The parser takes that number as an argument and never reads it, so six questions for a twenty-five question request is reported as a success.
'Frequently asked' is a counting problem: read the question papers, count the repeats. The hard part is that 'nothing repeats' and 'we failed to read a single paper' both arrive as an empty list.
We went looking for the scheduler in our AI study timetable and there is not one. No overlap detection, no chronological sort, no feasibility check — the plan is whatever the model returns, and the rules exist only as sentences asking it to behave.
The Home card and the notification both need to know which item is today's. One of them runs in a background isolate that cannot open the database at all — so instead of syncing a cursor, we deleted it.
We asked a diffusion model for a labelled diagram and got one back with PISITON written on it. Every label a student reads in this feature is now painted by Flutter, and the image is told not to draw text at all.
Our AI returns its answer as JSON with LaTeX inside it. Nothing errored, and the formula still arrived destroyed — because \frac is valid JSON that decodes to a control character.
Our two-speaker audio overview highlights the line being spoken as it plays. The finished recording contains nothing that says where a turn begins — so the timings had to be measured while the audio was still being made.
Adding a marketplace to a study app is easy. Adding one that is allowed to exist is a different problem — reporting, blocking, review queues and a daily cap.
The VTU result server sends an incomplete certificate chain. Android then has to fetch the missing piece over plain HTTP — and blocks it. Here is how a misconfigured server becomes a bug in your app.
Our app opened, five background jobs woke up at once, and every one of them reached for the same API key. Here is the queue that fixed it — and why the chat feature is deliberately not allowed to use it.
We could have filled in your credits automatically and it would have looked brilliant. Here is why we left the fields empty, and why the calculator rejects two whole schemes on purpose.
Tick twenty topics on a bad connection, close the app, and it still finishes the job. Here is the retry queue that made AI generation durable.
The search was working perfectly. It just never found anything — because truncated embeddings are not unit vectors, and the threshold assumed they were.
