Technical writing on dialogue systems, NLP architecture, and conversational AI from the Equmenopolis engineering team.
Most dialogue systems skip proper coreference resolution and pay for it in fallback rates. We tracked 6,000 production turns and found three root causes that all trace back to one architectural decision.
Read article
Flat intent lists work until users start combining requests. "Book a flight and cancel my hotel" is not two separate utterances - it's a compound intent with shared context. Here's how we represent that as a typed graph.
Read article
Context management has a latency budget. If state operations consume 150ms, you have 50ms left for inference - which forces you to use smaller, less capable models. We describe the state engine design that keeps overhead below 10ms.
Read article
Most teams track one dialogue metric: containment rate. It's easy to measure and largely useless for diagnosing quality problems. Here are the six metrics that tell you where your bot is actually failing.
Read article
Rasa and Equmenopolis take fundamentally different approaches to multi-turn dialogue management. Here's where each wins and where each fails, based on production deployments in comparable customer service use cases.
Read article
Ask a product team running a support chatbot what their containment rate is, and most can tell you a number. Ask them how they define containment and the answers diverge widely - because they're measuring different things.
Read article
A user disconnects mid-flow. They reconnect. The bot greets them as if they've never spoken. This cold restart problem is entirely solvable with the right architecture - here's how persistent dialogue memory works.
Read article
NLU model drift does not announce itself. Accuracy can degrade from 94% to 87% over six months without a single error log changing. Here's how to detect it before it becomes a support problem.
Read article
We closed a $4.2M Seed Round led by Emergence Capital to scale our team and EU infrastructure. Here's what we're building with it and why context management matters more than model size.
Read article
Teams that route ASR output through a chat NLU pipeline consistently underestimate the rework required. A model at 94% accuracy on chat may drop to 79% on voice. Here's why and what to do about it.
Read article
SOC 2 Type II is necessary but not sufficient for evaluating a conversational AI vendor. Here are the questions specific to dialogue session data that SOC 2 does not answer.
Read article
The translation pipeline approach - translate to English, run your NLU, translate back - fails in three specific ways. Teams that try it consistently rebuild with per-language models within 12 months.
Read article
A B2B SaaS platform was generating 40% of their support tickets from onboarding abandonment. Here's the architecture of the fix, what failed during deployment, and what the 90-day data showed.
Read article
Slot schema design errors are the most common cause of multi-turn dialogue failures after coreference problems. Here's how to design schemas that handle partial fills, conditional slots, and conflicting values correctly.
Read article
When a user escalates from a bot to a live agent, two things can happen. In one version, the agent already knows everything. In the other, the agent asks the user to repeat themselves. The difference is context transfer design.
Read article