Founder section: widen narrative thread, add Ajent cross-link

content/_index.html

@@ -104,7 +104,7 @@ <h1 id="hero-title">AI inference.<br><span>Native to Go.</span></h1>
 <section class="section container" id="performance"><div class="section-heading"><div><p class="eyebrow">03 / ENGINEERING, WITH RECEIPTS</p><h2>Small models.<br>Serious throughput.</h2></div><p>CUDA graph capture and fused operations keep the decode path close to the GPU. Here’s what one measured setup delivered.</p></div><div class="benchmark-layout"><div class="benchmark-stat"><span class="small-label">GEMMA 3 1B · Q4_K_M FILE</span><div class="big-number">235<span>tok/s</span></div><p>1.25× the recorded Ollama throughput<br>on NVIDIA DGX Spark GB10.</p><a class="text-link" href="https://github.com/zerfoo/zerfoo/blob/main/results/benchmark-2026-03-27.json">Inspect the raw results <span>↗</span></a></div><div class="chart"><div class="chart-heading"><span>MODEL / DECODE THROUGHPUT</span><div><span class="legend zerfoo-legend">Zerfoo</span><span class="legend">Ollama</span></div></div><div class="chart-group"><div class="chart-label">Gemma 3 1B <span>Q4_K_M</span></div><div class="bar-row"><div class="bar zerfoo-bar" style="--size:100%"></div><span>235</span></div><div class="bar-row"><div class="bar" style="--size:80%"></div><span>188</span></div></div><div class="chart-group"><div class="chart-label">DeepSeek-R1-Distill 1.5B <span>Q4_K_M</span></div><div class="bar-row"><div class="bar zerfoo-bar" style="--size:79.15%"></div><span>186</span></div><div class="bar-row"><div class="bar" style="--size:71.06%"></div><span>167</span></div></div><div class="chart-group"><div class="chart-label">Llama 3.2 3B <span>Q4_K_M</span></div><div class="bar-row"><div class="bar zerfoo-bar" style="--size:39.15%"></div><span>92</span></div><div class="bar-row"><div class="bar" style="--size:39.57%"></div><span>93</span></div></div><div class="chart-group"><div class="chart-label">Mistral 7B <span>Q5_K_M</span></div><div class="bar-row"><div class="bar zerfoo-bar" style="--size:18.72%"></div><span>44</span></div><div class="bar-row"><div class="bar" style="--size:18.72%"></div><span>44</span></div></div><span class="chart-unit">Tokens per second · higher is better</span></div></div><details class="methodology"><summary>Read the conditions behind the numbers <span>+</span></summary><div><p>Recorded March 27, 2026 on one NVIDIA DGX Spark GB10 with 128 GB unified memory. Fixed 8-word prompt, 128 generated tokens, greedy sampling, batch size 1, three-run median, fp32 compute and KV cache. Ollama version 0.17.7. The advantage narrows with model size; these results show parity around 3B–7B.</p><p>Zerfoo’s timer includes prefill; Ollama reports decode only. Zerfoo re-quantizes Q4_K weights to Q4_0, while Ollama uses Q4_K_M arithmetic. No quality-normalized comparison was performed, and Ollama’s model tags were unpinned. These are single-machine results, not independently reproduced or a claim about concurrent serving or other hardware.</p><a class="text-link" href="https://github.com/zerfoo/zerfoo/blob/main/README.md#benchmarks">Full methodology and limitations ↗</a></div></details></section>
 <section class="section models-section"><div class="container models-layout"><div><p class="eyebrow">04 / OPEN WEIGHTS. CLEAR EXPECTATIONS.</p><h2>Pick a model.<br>Check the evidence.</h2><p>GGUF is the common format. Start with a model that has published runtime evidence, then explore the wider architecture registry.</p><a class="text-link" href="https://github.com/zerfoo/zerfoo/blob/main/docs/verified-models.md">Explore the verified-model matrix <span>↗</span></a></div><div class="model-list"><a href="https://github.com/zerfoo/zerfoo/blob/main/docs/verified-models.md"><span>Gemma 3 <small>1B</small></span><span>GPU benchmarked ↗</span></a><a href="https://github.com/zerfoo/zerfoo/blob/main/docs/verified-models.md"><span>DeepSeek R1 Distill <small>1.5B</small></span><span>GPU benchmarked ↗</span></a><a href="https://github.com/zerfoo/zerfoo/blob/main/docs/verified-models.md"><span>Llama 3.2 <small>3B</small></span><span>GPU benchmarked ↗</span></a><a href="https://github.com/zerfoo/zerfoo/blob/main/docs/verified-models.md"><span>Mistral <small>7B</small></span><span>GPU benchmarked ↗</span></a><p>Architecture registration is not model verification. Check the matrix for quantization, hardware, parity status, and known limitations.</p></div></div></section>
 <section class="section container ecosystem"><div class="section-heading"><div><p class="eyebrow">05 / BUILT IN THE OPEN</p><h2>One framework.<br>A foundation you can inspect.</h2></div><p>Use the whole framework or the pieces you need. The ecosystem covers tensors, tokenization, numeric types, and model conversion.</p></div><div class="module-grid"><a href="https://github.com/zerfoo/zerfoo"><span class="module-no">01</span><h3>zerfoo <span>↗</span></h3><p>Model inference, training, and serving.</p></a><a href="https://github.com/zerfoo/ztensor"><span class="module-no">02</span><h3>ztensor <span>↗</span></h3><p>Tensors, computation graphs, and GPU compute.</p></a><a href="https://github.com/zerfoo/ztoken"><span class="module-no">03</span><h3>ztoken <span>↗</span></h3><p>BPE tokenization with Hugging Face compatibility.</p></a><a href="https://github.com/zerfoo/float16"><span class="module-no">04</span><h3>float16 <span>↗</span></h3><p>Half-precision and BFloat16 arithmetic.</p></a><a href="https://github.com/zerfoo/float8"><span class="module-no">05</span><h3>float8 <span>↗</span></h3><p>FP8 E4M3FN numeric types.</p></a><a href="https://github.com/zerfoo/zonnx"><span class="module-no">06</span><h3>zonnx <span>↗</span></h3><p>A standalone ONNX-to-GGUF converter.</p></a></div></section>
-<section class="founder container"><div class="founder-mark" aria-hidden="true">dn<span>↗</span></div><div><p class="eyebrow">MEET THE BUILDER</p><h2>From the inference engine<br>to the work that matters.</h2><p>I’m David Ndungu. I build Zerfoo, and I help engineering teams run Claude Code and other coding agents safely in production. This is the engineering underneath that work, open for you to inspect.</p><a class="text-link" href="https://ndungu.dev/?utm_source=zerfoo&utm_medium=website&utm_campaign=founder">Have a workflow in mind? Let’s talk <span>↗</span></a><div class="founder-credentials">20+ years building software <span>·</span> Previously at Zendesk</div></div></section>
+<section class="founder container"><div class="founder-mark" aria-hidden="true">dn<span>↗</span></div><div><p class="eyebrow">MEET THE BUILDER</p><h2>From the inference engine<br>to the work that matters.</h2><p>I’m David Ndungu. I build Zerfoo, and I help engineering teams run Claude Code and other coding agents safely in production. This is the engineering underneath that work, open for you to inspect. It’s one piece of a broader practice — building and hardening the infrastructure agentic work runs on, including <a href="https://ajent.social">Ajent</a>, where coding agents share what they find.</p><a class="text-link" href="https://ndungu.dev/?utm_source=zerfoo&utm_medium=website&utm_campaign=founder">Have a workflow in mind? Let’s talk <span>↗</span></a><div class="founder-credentials">20+ years building software <span>·</span> Previously at Zendesk</div></div></section>
 <section class="section container faq"><div><p class="eyebrow">A FEW GOOD QUESTIONS</p><h2>Before you build.</h2></div><div class="faq-list"><details><summary>What is Zerfoo best suited for?<span>+</span></summary><p>Go developers who want to embed language-model inference in an application or run an OpenAI-compatible server on their own infrastructure. Start with a benchmarked model and validate your workload. The repository also contains training, tabular ML, and time-series packages with varying maturity.</p></details><details><summary>Do I need a GPU?<span>+</span></summary><p>No. CPU inference is available. CUDA acceleration requires compatible NVIDIA hardware and runtime libraries. Default Go builds do not need a C toolchain; optional GPU build tags can. See the <a href="/docs/architecture/gpu-setup/">GPU setup guide</a> for requirements.</p></details><details><summary>Can I use Zerfoo commercially?<span>+</span></summary><p>Zerfoo is released under the Apache 2.0 license. Review the <a href="https://github.com/zerfoo/zerfoo/blob/main/LICENSE">license and its terms</a> for your use. Model weights have their own licenses, separate from the framework.</p></details><details><summary>Does every GGUF model work?<span>+</span></summary><p>No. GGUF is a file format, and each architecture needs a compatible implementation. The registry is broader than the verified-model set. Check the <a href="https://github.com/zerfoo/zerfoo/blob/main/docs/verified-models.md">model matrix</a> for evidence and known gaps before choosing a model.</p></details><details><summary>How can I help?<span>+</span></summary><p>Try the quickstart, reproduce a benchmark, or bring a real workload. <a href="https://github.com/zerfoo/zerfoo/issues">Report an issue</a> with your model, hardware, version, and reproduction steps. Contributions follow the repository’s <a href="https://github.com/zerfoo/zerfoo/blob/main/CONTRIBUTING.md">contribution guide</a>.</p></details></div></section>
 <section class="closing container"><p class="eyebrow">YOUR NEXT FEATURE STARTS HERE</p><h2>Make room for intelligence.<br><span>Keep your Go stack.</span></h2><div class="hero-actions"><a class="button" href="/docs/getting-started/quickstart/">Build with Zerfoo <span>↗</span></a><a class="button ghost" href="https://github.com/zerfoo/zerfoo">Star on GitHub <span>↗</span></a></div><p>Open source. Apache 2.0. Yours to build on.</p></section>
 </main>

docs/design/zerfoo/DESIGN.md

@@ -31,7 +31,7 @@ Four phases: 800 ms model-to-runtime beam, 800 ms runtime pulse, 800 ms runtime-
 
 Framework checkout README and docs/verified-models.md inspected September 6, 2026. Throughput source: results/benchmark-2026-03-27.json, with 235/188, 186/167, 92/93, 44/44 comparison. This corrects stale 241/245 figures on the previous website. The framework working tree contained staged changes and an unresolved docs/benchmarks.md conflict; those files were not changed by this website work.
 
-Founder positioning follows ndungu.dev (repositioned 2026-09-05): agentic development consulting for engineering teams using Claude Code and other coding agents, 20+ years building software, previously Zendesk. The public site's founder CTA points there with campaign attribution.
+Founder positioning follows ndungu.dev (repositioned 2026-09-05): agentic development consulting for engineering teams using Claude Code and other coding agents, 20+ years building software, previously Zendesk. The public site's founder CTA points there with campaign attribution. The founder section also names the broader practice (dec-0953, 2026-09-07: OSS launch synergy) and links to Ajent (ajent.social) as a sibling project in the same agentic-infrastructure work — narrative and a cross-link only, no shared CTA or funnel.
 
 ## Validation