Grading the model: the scorer hierarchy and evaluation pipeline

Inside NeuroLink's evaluation system: the BaseScorer/BaseLLMScorer/BaseRuleScorer class hierarchy, the PipelineBuilder-driven EvaluationPipeline that runs scorers together, and how batching, factories, and observability hooks fit around it.

Jun 23, 2026 Deep Dive, Engineering

Why Every Native Provider Must Wire the Same Tool-Persistence Hook

How a missing handleToolExecutionStorage call in a native Google Vertex code path can silently drop tool-call history from an agent's memory, and why every native provider must route through the same TelemetryHandler-backed persistence hook.

Jun 22, 2026 Engineering

Inside ConversationMemoryFactory: How NeuroLink Picks and Wires a Memory Backend

How NeuroLink's ConversationMemoryFactory picks between an in-memory store and Redis at startup, what the shared IConversationMemoryManager contract guarantees, and how the Redis backend avoids a tool-call/tool-result race condition.

Jun 21, 2026 Deep Dive, Engineering

ModelPool's error-class fallback design

How NeuroLink's ModelPool classifies a provider error into one of six classes before deciding whether a failed member gets a one-minute cooldown or is retired for the life of the instance.

Jun 20, 2026 Deep Dive, Engineering

Debugging WebRTC audio issues in production

A field guide to the five documented failure modes in NeuroLink's LiveKit voice agent, the STT/TTS credential check behind "no assistant audio", and the per-Job process isolation and inactivity watchdog that explain two common false alarms before you reach for those five checks.

Jun 12, 2026 Engineering, Tutorial

Scaling LiveKit rooms for production voice agents

LiveKit Agents dispatches one Job per room to its own OS process, so NeuroLink's LiveKit voice agent scales concurrent calls by adding worker replicas rather than by multiplexing calls inside a single event loop.

Jun 11, 2026 Engineering, Deep Dive

Twenty-four providers, one BaseProvider: the adapter catalog

How NeuroLink's `BaseProvider` abstract class unifies OpenAI, Anthropic, SageMaker, and the rest of the provider catalog behind one contract for chat, embeddings, and image generation.

Jun 7, 2026 Deep Dive, Engineering

Flagging call-graph claims automatically

A deterministic gate that scans a draft for call-graph claims built on real symbol names but invented wiring -- "X calls Y", exact return types, a private/public label, a mermaid diagram of functions that never ran in that order -- and why it flags instead of verifying, pairing with a new drafter rule that tells the model to describe responsibility, not wiring.

Jun 3, 2026 Engineering, Deep Dive

From User Input to Provider API: The Five-Stage Message Flow

How NeuroLink turns a raw user message — text, file paths, URLs, or image buffers — into a provider-ready request through five stages: normalization, file processing, multimodal assembly, budget enforcement, and provider adaptation.

Jun 1, 2026 Deep Dive, Engineering

Seventeen file processors, six categories, one priority system

How NeuroLink's BaseFileProcessor lifecycle and priority-based ProcessorRegistry route seventeen file processors across document, data, markup, code, media, and archive formats.