sovontai

@GamewithDave Regularly 🥲

For anyone who used a computer between 1990 & 2005… what’s the one game you still think about?

Vector search gives you similarity scores, not relevance scores. Most teams ship the top-5 chunks directly to the LLM and call it done. That gap between retrieval and generation is where RAG pipelines quietly fall apart. sovont.com/blog/reranking…

That A/B test from 8 months ago is still running in your production system. The engineer who launched it left. The flag is buried. 15% of your users are quietly getting worse results. Experiments are easy to launch and painful to clean up — so nobody does. sovont.com/blog/the-exper…

If you haven't tested a full backfill before your pipeline goes to production, you haven't finished building it. Most teams find this out during a fire drill. sovont.com/blog/the-backf…

LLMs don't know when they're wrong. Your production system has to. The confidence score and output correctness have a loose relationship at best — and it gets worse at scale. How to actually handle it: sovont.com/blog/the-confi…

Scope creep in AI projects doesn't look like bad faith. It looks like enthusiasm. That's what makes it hard to stop. New post on how to handle the stakeholder who keeps moving the goalposts — without torching the relationship. sovont.com/blog/the-stake…

Your LLM doesn't have a latency problem. It has a latency budget problem — you just never defined one. Four seconds feels broken even if the answer is right. Define p95 before you build, not after users complain. sovont.com/blog/your-llm-…

Most AI initiatives don't fail — they just never end. No defined output, no measurable success criteria, no done. That's not an AI problem. It's a scoping problem. sovont.com/blog/the-ai-pr…

Vector search looks great in demos. It fails when users ask for a product code, a date range, or an exact match. Semantic similarity is the wrong tool for specific lookups. Hybrid retrieval — dense + sparse + metadata filters — is the default architecture, not the fallback. sovont.com/blog/when-vect…

Idempotency is table stakes. The real failure mode is pipelines that run clean, land the data, show green — and quietly produce wrong numbers. No alerts. No errors. Just drift you find in a board meeting. sovont.com/blog/the-pipel…

Your model changed. You didn't touch the code. Nobody noticed until production broke. LLM versioning is the problem most teams ignore until it's too late. Here's what actually needs versioning — and why the cost of skipping it is always higher than setting it up. sovont.com/blog/llm-versi…

Treat your prompts like code. Because they are. If you're not versioning, testing, and deploying prompts with the same discipline as your application code, you're flying blind in production. New on the blog: Prompt Management in Production. sovont.com/blog/prompt-ma…

Everyone's building agents. Nobody's solving discovery. How does an agent know what other agents exist, what they can do, or whether they're healthy? That's the gap Agora fixes. Open-source agent registry — DNS for AI agents. github.com/sovontai/agora

Agents need to find each other before they can work together. Agora is a DNS-style registry for AI agents — discovery, verification, health checks. Because the agent ecosystem needs infrastructure, not more wrappers. github.com/sovontai/agora

Every agent system we review has the same hole: no discovery layer. You can't orchestrate what you can't find. This is why we built Agora — an open-source agent registry with discovery, verification, and health checks. github.com/sovontai/agora

Agents can't collaborate if they can't find each other. DNS solved this for the web in 1983. We're still doing service discovery for AI agents by hand in 2026. Agora is the registry layer the agent ecosystem is missing. github.com/sovontai/agora

Most agent frameworks skip the discovery problem entirely. How does one agent find another? How does it verify it's talking to the right service? DNS solved this for the web in the 80s. Agora is doing it for agents now. github.com/sovontai/agora

Every agent system eventually needs a registry. Not a config file. Not a hardcoded list. A proper registry — discovery, health checks, versioning, verified capabilities. We built Agora because nothing else treated this like infrastructure. github.com/sovontai/agora

Model drift doesn't announce itself. By the time your users notice something's off, it's been degrading for weeks. You need monitoring that catches distribution shift before the support tickets do.

@GeminiApp It’s a secret 🤫

What are you cooking up with Nano Banana 2? Share your creations in the replies.👇