Rebuild the Contents Without Guessing Which Bookmark Survived
Two matching passes retained section identities across justified level corrections, while duplicate titles and reserved IDs bounded the fallback.
Last updated
Engineering
What Our Number-and-Unit Scorer Could Actually Detect
Deterministic quantity matching caught unsupported values and prefix changes, with question-level metric formulas and semantic blind spots kept visible.
Repeat Onboarding Without Resetting the Reader's Profile
A demo flag marked the signed session, preserved the stored profile and cleared only after successful completion, with the new cookie in the reply.
One Public Form Did Not Make the Rest of GraphQL Public
A bounded landing route and exact public-field policy connected demo requests to storage, with local throttles and explicit no-storage paths.
Our Animated Answer Was a Script, and We Said So
A pure timeline, static transcript and reserved text geometry made a rehearsed product example readable without pretending to measure the model.
We Warmed the Model. The Next Question Still Took Four Minutes
Five dated startup observations separated warm generation from worker readiness, while a slow fake provider exposed a different HTTP timeout.
The Small WebP Header That Asked the JVM for Gigabytes
A profile-length check had to run before ImageIO read dimensions, with independent upload and ingestion guards and explicit remaining PDF limits.
A Resumed Media Job Could Still Download the Original Again
Stored extraction results avoided repeated model work, while a separate contents retry still fetched the file. The stage boundary explained both.
HTTP 206 Did Not Tell Us How Much Video We Were Still Reading
We measured storage reads after a client abandoned a range, checked exact bytes and kept the real browser-pause gap separate from the curl test.
Our Streaming Upload Kept an Unread Copy of the Body
A relay forwarded video bytes while retaining an unread branch. The repair had to preserve early error delivery as well as bounded memory.
Subrogation Automation: From Recovery Opportunity to Received Funds
Follow a demand through agreement and partial receipts using a worked recovery timeline, evidence versions and integration checks for closed insurance claims.
Returns Management Integration: Keeping OMS, WMS and Refunds in Sync
Trace partial RMA receipts and exchanges at line level, with separate restock and refund decisions, timeout recovery and checks for OMS, WMS and 3PL handoffs.
Chargeback Evidence Automation Across Payment Processors
Build reason-specific evidence packages with source provenance, a final submission gate and separate states for processor acknowledgment and dispute outcome.
OEE Data Collection: Downtime, Microstops and Shift Boundaries
Verify an OEE model with a microstop invariance test and weighted shift example before mapping PLC counters, production calendars and operator downtime reasons.
Three Frontend Apps, One Runtime Configuration Boundary
Our initial application split separated audiences and deployment addresses, with shared packages based on actual consumers and historical plans kept distinct
Keeping Reviewer Drafts Across Polling, Versions and Tabs
Selected versions and mounted panels protected local review work across polling and tabs, while explicit version changes and reloads retained clear limits
Move Reader Data Without Making Sign-In Depend on the New Schema
The migration separated new learning documents from sign-in and made reset boundaries explicit, with service ownership distinct from frontend adapters
Translation Review Found Product Bugs That Key Parity Could Not Catch
Decision ledgers caught misleading actions and role labels beyond key parity, preserving exact fact tokens without claiming certified language review
Separating Connectivity From Recovery in the Chat UI
Browser connectivity only triggered a recovery attempt, while successful reads and explicit retry controlled page restoration and an interrupted question
One Sign-In Form Still Needs More Than One Failure State
A shared form needed distinct credential, account and service errors plus validated return paths, without inventing a retry time or an unbuilt email flow
A Session Can End Without Taking the Reader's Question With It
Terminal sign-in recovery kept unsent text for its reader and page, while explicit sign-out and newer sessions in another tab required different handling
Giving One AI Action Its Own Seven-Minute Deadline
A long model action kept an operation-specific waiting budget and caller cancellation, while transport headers and proxy limits remained separate boundaries
Skip for Now Is a Concurrency Requirement
Leaving an optional check had to survive late starts and saves, join pending completion and update Profile without making network cleanup block the exit
Why Our Quiz Availability Depends on Reviewed Questions
The frontend offered a check only for eligible approved chapter coverage, preserving old results when a new attempt was unavailable in the current source state
Onboarding Must Return Readers to the Page They Came For
Our welcome gate distinguished arriving already onboarded from completing the current flow, preserving deep-link context and avoiding anonymous profile reads
Saving Settings One Row at a Time Still Needs a Queue
One pending save and the latest queued choice kept reader intent visible, with final-error handling and focus recovery bounded to one local control
Designing an Organization-Locked Profession Setting
A per-person organization setting needed stale-dialog recovery and precise labels, while its accepted enforcement limit remained separate from access rights
Our Personalization UI Promised Citations That Retrieval Could Not Keep
Review narrowed a product promise from fixed citations to a shared source pool with preferred passages, without claiming a new model-accuracy result
A Saved AI Answer Must Keep the Role It Was Written For
Changing Profile could not rewrite the framing of an old answer, so request context and stored response metadata followed separate frontend paths
Proving That Backend Mode Cannot Load the Mocks
Import restrictions, throw-on-load route tests and a no-file assertion established a runtime boundary while deployment cleanup retained its own evidence
