Articles, one at a time.
Every piece cites primary sources and adds something you can act on: a worked example, a tested command, a documented limitation. Drafts that only summarize what already exists stay unpublished.
Google A2A 1.0 in Production: An Interoperability Audit
We tested Google A2A 1.0 between local Python agents: discovery, task lifecycle, artifacts, auth, pre-1.0 agent card migration, and where adapters remain.
Read →
Langfuse vs Arize Phoenix Self-Hosted: Traces, Evals, and Ops Cost
We ran Langfuse and Arize Phoenix locally, sent synthetic LLM traces over OpenTelemetry, exercised evals, and measured the operational cost of self-hosting.
Read →
Pydantic AI in Production: Type-Safe Agents vs LangGraph Overhead
We ran Pydantic AI's structured outputs, dependency injection, retries, and tool calls, then compared it with LangGraph across a 160-scenario test matrix.
Read →
DSPy GEPA vs Manual Prompts: Cost, Overfitting, and Model Upgrades
We benchmarked DSPy GEPA against manually engineered prompts, measuring held-out accuracy, optimization token cost, latency, and transfer to a newer model.
Read →
Windmill vs Airflow: A Hands-On Production Review for AI Workflows
We deployed Windmill with Docker Compose, built a mixed-language data and AI workflow, tested worker isolation, and mapped trade-offs against Airflow.
Read →
CodeRabbit vs Greptile: PR Review Catch Rate, Noise, and Cost
We tested CodeRabbit and Greptile on seeded bugs in TypeScript, Python, and Go pull requests: what each caught, where noise built up, and what seats cost.
Read →
FastMCP Streamable HTTP in Production: Auth, DNS Rebinding, Docker
We deployed FastMCP behind Docker and nginx with OAuth token introspection, then tested Streamable HTTP sessions, proxy headers, and DNS rebinding defenses.
Read →
SGLang RadixAttention vs vLLM on One H100: Throughput Reality Check
We benchmarked SGLang RadixAttention against vLLM prefix caching on one H100: shared prefixes, multi-turn chat, structured output, latency, and GPU cost.
Read →
Unsloth on One GPU: Llama 3.1 8B Throughput, VRAM, and Quality
We compared Unsloth with a standard Hugging Face TRL QLoRA pipeline on one RTX 4090: training time, VRAM, quality, dependency failures, and limits.
Read →
Pipecat vs LiveKit Agents: Voice Latency and Lock-In Benchmark
We benchmarked Pipecat, LiveKit Agents, and Pipecat over LiveKit with synthetic calls and identical voice providers: where latency spikes and what breaks.
Read →
E2B Sandbox in Production: Cold Starts, Costs, and Pause/Resume
We tested E2B sandbox lifecycle under multi-turn agent workloads: cold starts, per-second costs, auto-pause, resume failures, and parallel creation storms.
Read →
Docling in Production RAG: Where PDF Chunking and Tables Break
We tested Docling 2.92.0 on digital and scanned PDFs for table fidelity, chunk quality, memory pressure, and serialization failures in a RAG pipeline.
Read →