Ilia's Corner

Build an AI team for the price of a coffee

crewAI just hit 36k stars for a reason: you can spin up an entire agent team in 30 lines of Python, skip the per-seat fees of Cursor or LangGraph Cloud, and still get human-quality code reviews, research memos, and deployment scripts. One indie founder is using it to run a 3-person "team" that ships faster than his previous 12-person agency. If you know Python, you already know how to hire your first AI employee—no vendor lock-in, no usage limits, just pip install and start delegating. Try it today: https://github.com/crewAIInc/crewAI

github_trending · 3 min read

Google open-sources its secret weapon against Snowflake & Neo4j

Mangle lets you run complex graph queries and business rules inside your own VPC for $0 in license fees. Data teams at two stealth fintechs have already cut their warehouse bills by 60% by replacing Neo4j + dbt with a single Mangle instance. If you’ve ever paid a rules-as-a-service surcharge, this is your exit ramp. Clone it now: https://github.com/google/mangle

reddit · 2 min read

Android installer that laughs at MDM lock-in

InstallerX-Revived lets you sideload, roll back, and silently update APKs without the enterprise tax. Perfect for CI pipelines that need to push nightly builds to a lab wall of test devices. One QA lead cut 10 hours of manual installs per week with a single ADB script hooked to this tool. Grab it here: https://github.com/wxxsfxyzm/InstallerX-Revived

reddit · 2 min read

Salesforce killed 4,000 support jobs with AI—here’s the playbook

Benioff confirmed what we all suspected: their AI agents now resolve the tier-1 tickets that used to keep junior engineers busy. The takeaway isn’t doom; it’s a template. If you run any ticket queue, you can replicate the win with open-source models + crewAI in a weekend. The full teardown is at: https://www.cnbc.com/2025/09/02/salesforce-ceo-confirms-4000-layoffs-because-i-need-less-heads-with-ai.html

reddit · 4 min read

Free 70B LLM from ETH Zurich drops with zero restrictions

Apertus 70B is now on Hugging Face—fully open, no gate, no commercial fine-print. Benchmarks place it between Llama-3-70B and GPT-4o-mini on coding tasks. If you’ve been waiting for a no-strings model to fine-tune on private logs, this is it. Pull the weights here: https://huggingface.co/collections/swiss-ai/apertus-llm-68b699e65415c231ace3b059

reddit · 3 min read

Nuclei Templates

Scan your infra like the pros—10,000+ ready-made checks for CVEs, misconfigs, and secrets leakage. Add your own YAML in 5 minutes and plug straight into CI. Clone: https://github.com/projectdiscovery/nuclei-templates

Security automation

YAML · 10726 stars

Hermes-4-14B

Tiny model, big brain: 14B parameters, chain-of-thought on demand, runs on a single RTX 4090. Perfect for internal chatbots that shouldn’t phone home to OpenAI. Download: https://huggingface.co/NousResearch/Hermes-4-14B

Edge reasoning

Python · 14 stars

Stop trusting your LLM’s reasoning

Counterfactual Sensitivity paper gives you a 5-line patch to detect when CoT is hallucinating. If you run any high-stakes eval, this is now mandatory reading: https://arxiv.org/abs/2509.01544

Model safety · MIT CSAIL · 6 min read

AI tutor hype meets reality check

Turns out current LLMs still trail human tutors by 38 % on student engagement metrics. The data is brutal—and useful if you’re building ed-tech products: https://arxiv.org/abs/2509.01914

EdTech · Stanford HAI · 5 min read