Company
Blog
Engineering deep-dives, product updates, and AI economics insights from the Azoth team.
Introducing the IO Inventory Connector
Azoth now supports asynchronous vision extraction and catalog matching for IO data pipelines. Every ingest trace can be automatically enriched with image descriptions and SKU-level matching — without blocking the fast path.
GitHub Marketplace integration is live
Install the Azoth GitHub App in two clicks. Regression suites now run automatically on every pull request and post results as GitHub Check Runs — pass/fail, latency delta, cost delta, all in your PR review.
How we separated the ingest control plane from the data plane
Our original ingest route did too much: auth, budget check, PII scrub, database insert — all blocking the response. We split it into a sub-millisecond control plane and an async worker. Here's how.
The hidden cost of over-engineered prompts
After analyzing 50M AI calls across Azoth workspaces, we found that 38% of input tokens carry zero semantic information. Filler phrases, redundant context, deprecated instructions. Here's what the data looks like.
Multi-provider routing: when to use DeepSeek vs GPT-4o
Our model matrix now covers 9 providers and 40+ models. Not every task needs the most capable model. We break down which routing strategies save the most money for common workloads.
How a Series B fintech cut their AI bill by 58% in 30 days
They had 12 agents, 3 providers, and no visibility into what each one actually cost. After connecting Azoth, the Arbiter rerouted 64% of their GPT-4o calls to cheaper models with no quality degradation.
Stay up to date — subscribe by email