About Blackwell Systems
How I work
Evidence over assertion
Most technical claims are hard to check. My work is the opposite: everything I put forward is something I can show you. That is the through-line from the code to the advisory work.
Built in the open
The infrastructure is public, the benchmarks ship with their methodology and logs, and the numbers reproduce.
Verified on your data
When I advise, the recommendation is held to the same bar: measured on your own systems, not asserted from a template.
Clear about the edges
I state what a finding rests on and what would change it. Neutral by default, and direct about what is worth building.
Selected work
The open infrastructure the AI stack is building on
Production-grade, open source, and backed by measured, reproducible numbers. This is the evidence behind the judgment.
Building AI Agents in Go (book)
A 776-page book that teaches how AI agents work by building one in Go with the standard library, from a 30-line loop to a production crash-triage agent: tool calling, MCP, evals across models, sandboxing, and crash-safe runs. $29 at golang-agents.com.
GCF
An AI-native wire format reverse-engineered from tokenizer data. 100% comprehension on every frontier model, 91.2% on complex code graphs where JSON drops to 54.1%, 50-92% fewer tokens, 43B+ lossless round-trips. Seven implementations, 580K+ downloads. Used by 28 projects outside Blackwell Systems, including Chrome DevTools MCP, OmniRoute, Speakeasy, and the Linux Foundation ODPS.
bide
Durable AI agents in Go, with side effects that fire at most once. One append-only journal derives four guarantees no other framework pairs in a single library: at-most-once side effects (measured against a fair crash benchmark), thousands of concurrent durable runs in one process with no cluster, a cryptographically verifiable RFC 6962 Merkle audit trail, and provably convergent shared state (gsm, machine-checked in Coq).
knowing
A content-addressed code intelligence engine that beats every competitor in the category with statistical proof: 3.79x codegraph, 6x GitNexus, 22x grep across 302 tasks and 17 repos. Supply-chain detection without executing code, at a 1.0% false-positive rate.
agent-lsp
A stateful MCP runtime over real language servers: 65 tools, 30 CI-verified languages, and a speculative execution engine that simulates edits in memory before committing. On the official MCP Registry and Glama (A-tier).
mcp-assert
The deterministic testing standard for MCP servers. 55,000+ downloads, 102 servers scanned, 34 upstream bugs found, including 9 crashes in Ant Group's mcp-server-chart, fixed upstream. Runs in CI across 39 of WYRE's MCP server repos.
polywave
A formally specified parallel-agent coordination protocol: 6 invariants, 48 execution rules, 7 roles, and a 4-5x measured speedup. knowing (94K lines) was built on it.
GCP emulator platform
Local Google Cloud APIs for development and CI. The Secret Manager emulator is the most widely adopted community solution, ranked #1 on Google, Bing, and DuckDuckGo, with 50K+ downloads and enterprise CI adoption.
Research
The theory under the tooling
Eleven self-published research papers (Zenodo DOIs), several with industry response. They work out the theory the tooling is built on, from tokenizer design and transformer attention to distributed convergence and memory reclamation. The proof of value is the inbound they generated, not review status. Full set with abstracts on blog.blackwell-systems.com/about.
Tokenizer-attention coupling
How BPE merge decisions permanently shape transformer internal organization. 43 tokenizers, controlled experiments, 3-738x better structured-data comprehension at zero natural-language cost. doi:10.5281/zenodo.20925910
Stranded attention
A previously undescribed failure mode: structural capacity the tokenizer permanently prevents every attention head from using. doi:10.5281/zenodo.21158886
Content-addressing as a primitive
Hierarchical Merkle trees over code-relationship edges as a query-optimization substrate. The theory behind knowing. doi:10.5281/zenodo.20342255
Background
Production scar tissue, not just theory
The open-source work shows the range; the operating experience is why the judgment holds up under real constraints.
Revenue-critical systems at scale
Backend for a global loyalty platform serving millions of members across 120 countries, the foundational layer nearly every engineering team consumed. 24/7 on-call.
Systems I designed and ran
A digital wallet across 5 currencies, a serverless promotion rules engine, and a real-time CDC pipeline that replaced full-table scans, processing in about 1.3 minutes regardless of dataset size.
Why it matters to you
The advisory read is practical, not academic. I have owned the outage, the migration, and the cost line, so build-versus-buy and vendor calls come from having lived the consequences.
The practice
The same judgment, applied to your stack
Three services, one method. The full breakdown, with how each engagement runs, is on the services page.
Fractional CTO
Embedded technical leadership: strategy, build-versus-buy, architecture, and cost.
Technical due diligence
Independent evaluation of a claim or vendor, delivered as a decision memo with economics and risk.
AI cost optimization
A full-stack audit that cuts the LLM bill without losing quality, every number measured on your data.
Want an outside technical read?
Tell me what you are weighing and I will give you an independent, evidence-first answer.