Yam Marcovic
@ymarcov
Solving LLM Agent Alignment @ http://Parlant.io @emcieco
This is huge! This fixes the universal problem with conversational agents: hallucinations! Every company that builds customer-facing agents struggles to align them with business rules and make them consistent. If you solve this, you'll become a hero overnight. Check out…
With MCP being all the rage, there are important, counter-intuitive caveats to know about when building agentic APIs. It's *not* the same game as building regular APIs, at all. Different, often completely opposite rules apply! From deep practical experience in building…
One Token to Fool LLM-as-a-Judge Watch out for this one, devs! Semantically empty tokens, like “Thought process:”, “Solution”, or even just a colon “:”, can consistently trick models into giving false positive rewards. Here are my notes:
A really interesting blog post on Conversational AI by Parlant.io parlant.io/blog/why-we-bu…
Instructing LLMs for Customer Interactions Is Harder Than It Seems... Ever tried explaining a task to someone, only to realize they took your instructions WAY too literally? Now, imagine that "someone" is an AI automating your customer interactions. "Offer a discount when it…
AI Agents fail 99% of the time when real customers show up. Parlant fixes this by replacing random LLM decisions with structured conversation rules that actually work. Works with OpenAI, Gemini, Claude 4 and Llama models. 100% Opensource.
A while back, I was sitting with technical experts from a large enterprise, discussing our open-source product (Parlant.io) and how it helps create fluent yet strictly compliant customer-facing AI agents, even at scale. Then came a surprise: “Can we use Parlant…
Conversational AI devs, today I'll share insights and hard-earned lessons from working with some of the largest organizations in the world that are trying to implement customer-facing AI agents. Let no one tell you it's easy. In my talk on miniCON today, you'll learn the unhyped…
Just wasted 30 mins trying to create a simple Next.js landing page using @Windsurf. All I said was essentially "build me a nice site," nothing complicated. The result: error after error in loops. Not a single successful page load, but so far it's opened up like 5 servers,…
Could AI agents finally stop making things up? 𝐏𝐚𝐫𝐥𝐚𝐧𝐭 might be the answer. (It's 100% open-source) Imagine being able to build customer-facing AI that stays on message without sounding like it's reading from a script. That's what Parlant is tackling ➟ the core…
We just launched Parlant 2.0 🚀 In a nutshell, Parlant is a framework for building customer-facing GenAI agents, with a strong focus on customer experience and reliability. For more details github.com/emcie-co/parla…
ARQs are a better way for LLMs to reason on domain-specific problems. ARQs were born out of necessity in our development of Parlant (github.com/emcie-co/parla…) as, above all else, Parlant's goal is to achieve better results from LLMs in instruction-following accuracy and…
Just in case anyone's been losing sleep over this... —not. What's happening here is the industry's in a slump, so instead of innovating with meaningful abilities, we're starting to innovate with different ways of chewing up the same ideas. Empty calories's all it is. Fellow…
