Blogs
Writing on LLM testing, agent harnesses, and how transformers work, from the goML blog, Medium, and my own notes.
2026
- Claude Opus 5.5 LLM testinggoML blogGoML’s Opus 5.5 LLM testing covers benchmark results, critical bug fixes, missed tasks, cost, safety and deployment decisions for enterprise coding.
- Grok 4.7 LLM testing v/s enterprise realitygoML blogGrok 4.7 LLM testing covers coding performance, speed, cost, data handling, and AI Matic Bench Score results for enterprise AI workloads.
- Harness Engineering for AI AgentsgoML blogThe missing layer for production deployment of AI agents.
- LLM Testing of Muse Spark 1.3goML blogPutting Muse Spark 1.3 through GoML's evaluation suite to see where it lands.
- GPT 6 LLM Testing: What the Benchmarks Mean for Enterprise DeliverygoML blogWhat the GPT-6 benchmarks mean for enterprise AI delivery.
- LLM Testing of Anthropic's Hacker-Opus TestgoML blogHow Anthropic's model performs on the hacker-opus test.
- LLM Testing of Gemini 3.8 Flash: Scored on the AI Matic BenchgoML blogBenchmark results for Gemini 3.8 Flash on the AI-matic bench.
- The Missing Layer Between Your Agent and ProductionMediumHarness engineering: boundaries, repair loops, and the verification gate that decides what ships.
- The Part of Frontend the LLMs Can't DoMediumSkills, component libraries, shader kits, icon sets, inspiration galleries, and design craft.
- AI Agent Evaluation for Self-Improving Agent LoopsgoML blogAn evaluation framework for autonomous agents that iterate and improve on their own execution loops.
- Claude Watermarking and Enterprise AI TeamsgoML blogWhat Claude's watermarking approach changes for enterprise AI teams and provenance verification.
- A Practical Guide to Graph Engineering by GoMLgoML blogGoML's practical guide to graph engineering, knowledge graphs, and agentic traversal patterns.
- LLM Testing of Grok 4.6goML blogGrok 4.6 benchmarked, framed as a cost-curve shift rather than a capability jump.
- The Ratio of Quiet ThingsNoteA good bookmark is a promise you make to a future self who has different interests and less time.
- I Made My API 3,400x Faster, Then I Felt BadMediumA read-through cache story about Memcached, Neon Postgres, and the seductive lie of a big number.
- LLM Testing of Claude Opus 5goML blogTesting Claude Opus 5 as the first frontier model GoML calls enterprise-ready.
- Nine MinutesNoteThe spiral took four evenings. The figure, nine minutes.
- Debugging as PrayerNoteBoth begin the same way: something is wrong and I do not know what.
- LLM Testing of OpenAI GPT-5.6goML blogEvaluating OpenAI's GPT-5.6 variants, Sol, Terra, and Luna across reasoning and latency.
- Grok 4.5 (High) Model OverviewgoML blogAn overview and internal evaluation of Grok 4.5 (high) performance and tooling integration.
- Instructions for a StreetlampNoteStand where they left you. Let the moths mistake you for the moon.
- The Complete Guide to Sonnet 5goML blogGoML's complete guide to Sonnet 5 architectures, context utilization, and latency optimization.
- Sakana AI Fugu Enables One API for Production AIgoML blogHow Sakana AI's Fugu enables one unified API for smarter model routing and resilient production AI.
- The Harness Is the ProductMediumWhy the orchestration harness matters more than the raw model.
- Rogue Agent Impact VisualizergoML blogA visualizer and blast-radius calculator for detecting and mitigating rogue AI agent behaviors.
- Cutting a 3-Hour AWS Investigation to 11 MinutesgoML blogCutting a 3-hour AWS observability incident investigation down to 11 minutes with automated telemetry agents.
- The Complete Guide to AWS DevOps AgentgoML blogGoML's complete guide to configuring, deploying, and hardening the AWS DevOps Agent.
2025
- Amazon Bedrock Endpoints ExplainedMediumA practical breakdown of Bedrock endpoint types, invocation patterns, and how to choose the right setup.
- How Transformers Generate TextMediumHow decoding strategies (greedy search, top-k, nucleus sampling, and temperature) shape model output.
- The Many Minds of TransformersMediumWhat each attention head is really learning, and why running several in parallel gives richer representations.
- Attention Is All You Need to UnderstandMediumThe scaled dot-product attention mechanism demystified: queries, keys, values, and the matrix math.
- Transformers Explained SimplyMediumStart from raw tokens, walk through embeddings and positional encoding, and arrive at self-attention.
- The Only Prerequisites You Need for NLPMediumThe exact math, Python, and ML fundamentals you need before diving into language models.