Ilia's Corner

Stop paying Adobe and Mathpix—open-source now converts any file to LLM-ready Markdown

Microsoft's MarkItDown and Datalab's Marker just killed the last excuse for pricey document-conversion APIs. Drop a 200-page PDF, a PowerPoint deck, or even a YouTube link into either tool and seconds later you get clean Markdown ready for vector stores, fine-tuning prompts, or RAG pipelines—no Azure credits, no per-page fees. Teams are already piping the output straight into LangChain and LlamaIndex, cutting preprocessing costs to zero while keeping full data privacy.

GitHub · 3 min read

Google Wire rips Spring and Dagger out of your Go services

Trade annotation magic for 100-line Go files and watch cold-start latency drop 30-40 %. Wire generates DI glue at compile time, so you ship a single static binary with zero reflection or framework bloat. Perfect for serverless functions where every millisecond costs money.

reddit · 2 min read

Magentic-UI turns any LLM into a browser puppeteer

Need to automate a legacy web portal that has no API? Spin up the container, feed the model a prompt like 'download last quarter’s invoices,' and watch it click, scroll, and fill forms for you. Great for ETL teams stuck with browser-only SaaS dashboards.

reddit · 4 min read

SpatialLM builds BIM models from a phone video

Point your iPhone at a room for 30 seconds; get back a structured scene graph with walls, doors, and 59 furniture classes at centimeter accuracy. Architects are feeding the output straight into Revit, skipping days of manual drafting.

reddit · 3 min read

Zulip 11.0

Slack and Discord keep tightening APIs—Zulip just shipped threaded chat that you can self-host in five minutes with Docker. Same real-time feel, zero vendor lock-in, and you keep your data on your own Postgres cluster.

Team chat without lock-in

Python · 9 stars

JMAP MCP

Plug your email directly into AI agents using the open JMAP protocol. No paid Zapier tiers, no proprietary connectors—just a standards-based inbox your agents can read and write like any other MCP resource.

Agent email access

Go · 19 stars

Intel Core Ultra now lets you steal RAM for the iGPU

Flip one BIOS switch and your integrated graphics sees up to 50 GB of shared memory. That turns a $900 ultrabook into a credible local LLM runner—7 B models at 20 tokens/sec, no discrete GPU required.

Hardware · VideoCardz.com · 2 min read

Google Imagen 4 tiers launch in Gemini API

Need a hero image for a blog post? $0.02 and two seconds. Need a print-ready poster? Bump to Imagen 4 High and get 4 k fidelity. The same API call now lets you pick speed vs. quality on demand.

AI · Google Developers Blog · 3 min read