Engineering 63
- DeepSeek gotchas on NeuroLink
- How we built a 14-gate AI content pipeline
- Testing proxyIR for exhaustiveness
- Morph provider: quirks and workarounds
- Migrating a generate()-based judgment call to decide()
- Managing a 64-model catalog without chaos
- generate, stream, decide: a third inference type for NeuroLink
- When max_tokens truncates structured output, NeuroLink still said "stop"
- Correlating traces across providers
- Integrating GMI Cloud
- Cerebras integration deep dive
- What a green provider sweep proves, and what it skips
- The byte-cursor ledger: tracking proxy state precisely
- Designing the /accounts endpoint
- The Mistral quirk we had to special-case: registryDefaultModelChecksEnvVar
- Defending against decompression bombs
- Latency engineering for realtime voice agents
- Coordinating multiple agents without collisions
- Deterministic replay for proxy debugging
- Zero-downtime updates for the proxy
- Scraping the dependents graph to find real users
- What our first AI share-of-voice baseline actually measured
- The Gemini thoughtSignature bug
- Two backends, two stores, one TaskManager: how NeuroLink schedules AI work
- Embedding fast-path and session stickiness
- NeuroLink Brownfield Integration: Express & Fastify Adapters
- Voice as Three Stream Topologies: TTS, STT, and Full-Duplex Realtime
- Inside NeuroLink's RAG: The generate() Integration Bridge
- Grading the model: the scorer hierarchy and evaluation pipeline
- Why Every Native Provider Must Wire the Same Tool-Persistence Hook
- Inside ConversationMemoryFactory: How NeuroLink Picks and Wires a Memory Backend
- ModelPool's error-class fallback design
- Debugging WebRTC audio issues in production
- Scaling LiveKit rooms for production voice agents
- Twenty-four providers, one BaseProvider: the adapter catalog
- Flagging call-graph claims automatically
- From User Input to Provider API: The Five-Stage Message Flow
- Seventeen file processors, six categories, one priority system
- Four-stage context compaction: what runs when the model window fills up
- What You Actually Inherit When You Extend BaseProvider
- Symbol-grounding: catching hallucinated APIs before publish
- Ten ESLint rules that hold NeuroLink's type system together
- Disaster recovery from a session log
- Jina embeddings on NeuroLink
- How we shipped 12 providers in one PR
- xAI / Grok integration deep dive
- Fireworks vs Perplexity on NeuroLink
- Together AI deep dive
- Choosing between image, avatar, and music generation APIs
- Combining avatar and voice for talking-head output
- Hitting the SmolLM2 ceiling
- Rolling out 4 new providers (2 local, 2 cloud) — and the AI SDK bug we hit
- stepIndex bookkeeping: keeping multi-step calls in order
- How We Test NeuroLink: 20 Continuous Test Suites and Counting
- Meet Tara: From Slack Thread to Merged PR
- The hold-and-rewrite pipeline for stale drafts
- The silent Vertex default in your RAG pipeline
- How We Scaled to 13 Providers: The Provider Registry Story
- How We Built Multi-Provider Failover: Never Losing an API Call
- How We Built MCP Integration: Supporting 4 Transport Protocols
- How We Built the RAG Pipeline: 10 Chunking Strategies and Why
- How We Built Streaming Tool Calls: Real-Time AI at Scale
- Command injection in the ollama integration