Creează un cont pentru a te abona la acest blog și a-i urmări articolele pe pagina Lectură.
-
Two new analytics endpoints for aggregate cost data and per-job cost details, building on the token tracking work to answer 'how much has this thing cost me?'
-
Standardizing on KaTeX for all math rendering after the presentation modes feature broke LaTeX with inconsistent library choices and CSP blocking.
-
The biggest feature since launch — Native, Non-technical, and Kid-friendly modes with DynamoDB-stored prompts — plus accurate token tracking across both Haiku and Opus calls, and a waitlist for quota-exhausted users.
-
Adding a proactive daily quota system that disables uploads before users waste time, and fixing a bug where DynamoDB's TTL was silently deleting completed jobs.
-
Running deploy-test.sh for the first time and verifying the entire stack works end-to-end in the test account, from Secrets Manager to CloudFront.
-
Separating the single deploy script into test and prod versions with account guards, and setting up a fully independent test environment with its own S3 buckets, domain, and ACM certificate.
-
Another round of Content Security Policy fixes, two new operational API endpoints for monitoring, and a deep dive into why the screenshot Lambda was silently failing — plus the fix.
-
Adding Google Analytics across all pages and improving the daily upload limit error message to surface the actual limit value.
-
Replacing Claude's ad-hoc field classification with the NSF's NCSES Taxonomy of Disciplines, moving classification from Opus to Haiku to save costs, and adding a FAQ page.
-
The busiest day yet — a complete landing page overhaul, auto-scrolling carousel, field metadata, user-friendly errors, prompt injection defense, a full gallery page with screenshots, and two security regressions found and fixed.
-
Moving LLM configuration to SSM Parameter Store for zero-deploy changes, and replacing direct multipart uploads with pre-signed S3 URLs to handle large PDFs.
-
Extracting paper metadata from Claude's responses to build a browsable library carousel, and the bugs that surfaced once real data hit the frontend.
-
Adding financial guardrails with daily rate limits and per-job cost tracking, cleaning up the footer and secrets, and replacing GitHub Gist publishing with direct S3 uploads and a custom domain.
-
How NowIGetIt went from a simple idea — send a PDF to Claude, get back an interactive web page — to a working prototype with AWS deployment, streaming fixes, and a dark-themed UI, all in one day.