Hybrid local and cloud LLM stack for regulated financial document processing?I'm scoping a hybrid AI pipeline for a consulting client in a regulated industry (GLBA-covered, NPI involved). Trying to validate the architecture before bringing on an engineer to build it.The workflow: ingest financial PDFs (bank, brokerage, retirement statements, tax returns), classify by asset type, extract data, apply domain-specific business logic, populate Excel templates and fillable PDF forms. Compliance constraint: no NPI can hit a cloud API without ZDR-style controls.Current architecture sketch: - Local LLM (Ollama or LM Studio) on dedicated hardware for OCR and first-pass extraction - Local PII scrubber/tokenizer (Presidio or Skyflow) replaces identifiers with tokens before any cloud call - Cloud LLM under enterprise terms (Claude API with ZDR, or Bedrock equivalent) for the reasoning layer - Local de-tokenization and template populationQuestions for anyone who's actually shipped this pattern: 1. What stack did you land on, and what would you do differently? 2. Local model
Show HN: Tools.video – Browser-native video processing at 50 GB scaleHi HN, I built tools.video (https://tools.video) — 18 free video and audio tools that run entirely in your browser. Files never leave your machine.Compress, convert, trim, crop, resize, merge, rotate, reverse, speed change, screenshot, mute, video-to-GIF, audio extract/trim/convert/merge/speed, and media info.The interesting technical problem: FFmpeg compiled to WASM works, but it tops out around 4 GB, runs single-threaded, and is 10x slower than native browser APIs for the same operations. I wanted to handle 50 GB files and actually use the hardware (theoretically there's no limit at all).So I built a hybrid engine:- Most tools use WebCodecs (via MediaBunny) — the browser's native hardware-accelerated video encoder/decoder. H.264, HEVC, VP9, AV1 with GPU acceleration when available. This is what makes it fast. - Three tools (speed, merge, video-to-GIF) still use FFmpeg WASM where WebCodecs can't do the job. - The engine auto-detects which path to use and falls back gracefully — if har
Collaborating minds think alike, processing information in similar ways in a shared task, study showsWhether great minds think alike is up for debate, but the collaborating minds of two people working on a shared task process information alike, according to a study published in PLOS Biology by Denise ...
Microsoft Copilot Intros Voice Commands, Teams Collaboration, Local Data ProcessingMicrosoft has added new features within its Microsoft 365 Copilot offering, aimed at making further foothold in the enterprise, including voice-based interaction, group collaboration tools, and an ...
Show HN: AI Photo Cleaner iOS app (on-device processing)Hi HN,I've recently published my first iOS app to the Appstore! Its a photo cleanup app which uses on-device Apple Vision models to cluster similar images together into groups, and then ranks each photo within each group.Link: https://apps.apple.com/app/ai-photo-cleaner-sorter/id6754172...How it works: The best photo from each group gets marked as "Saved" by default, however in the Review stage, you can select others from the group to save if you want or unselect the "Best" one and choose another one to keep. The general concept is "keep the best, remove the rest", and in theory if you agree with the clustering/scoring and didn't want to make any amendments, you could clean up 100's of photos (and MBs) with just a few taps.Motivation: I'm one of those people who somehow builds up an insanely bloated camera roll - 15 photos of the same restaurant dish, 10 versions of the same sunset. So I initially built this for myself to help cleanup those 500 photos taken during a 1 week vacation...
Microsoft Expands Copilot With Voice Commands, Teams Collaboration and Local Data ProcessingMicrosoft is broadening its Microsoft 365 Copilot offering with new features aimed at making further foothold in the enterprise, including voice-based interaction, group collaboration tools and an ...