What I build
My GPU kernels are measured against AMD's own library under a frozen protocol. The receipt keeps the wins, the misses and the hardware state together. Decode and serving claims stay limited to the checks that actually ran.
-
GPU kernels, measured against the vendor's own library
Matrix kernels written in Mojo for RDNA3, run against AMD's hipBLASLt under a protocol written down before the numbers existed, including which results would count as a loss. The receipt publishes misses alongside wins.
LimitOne architecture (gfx1100), one vendor library, square fp16 GEMM plus one decode shape. Nothing here says anything about NVIDIA, about training, or about a full inference stack.
-
Android apps that keep working with the network off
Three shipped apps that move camera, files and documents between devices without a server in the path. One of them requests no internet permission at all — not as a policy, as a manifest. The privacy policy for each is a page on this site rather than a PDF, because a claim you cannot link to is not a claim.
LimitAll three are in closed testing on Google Play. Store listings are not public yet, so you cannot install them today — you can only read what they promise.
-
Checkout and payment systems, built to the boring standard
Merchant onboarding, hosted checkout, refunds, webhook delivery with a dead-letter path, and per-event projections — the parts of a payments product that are unglamorous and expensive to get wrong. Mail failures never fail the money path; that is a rule enforced in code, not a habit.
EvidenceNot published — this is client-shaped work, and it names other people's businesses
LimitI integrate a payment provider; I am not one. Card data never touches my servers, and nothing here is a claim to hold funds or to be a licensed institution.
-
The infrastructure under this page is mine, including the analytics
This site, its mail, and its visitor statistics run on hardware I administer. The analytics are a self-hosted build behind a consent gate that ships no script at all until you opt in — verified by a test that fails the build if a request escapes before consent. No third-party tag manager, no advertising network, no data broker.
LimitOne operator, one box, no high-availability story. If it goes down at 3am, it is down until I wake up. That is an acceptable trade for a portfolio and a deliberate one.
-
Software in eight languages, with a gate that fails on a missing string
Eight locales across five surfaces, English as the single source of truth, and a test that refuses the build if a key is referenced but untranslated, translated but never referenced, or present in one language and absent in another. No plural machinery in any language, because plural rules are where localisation quietly breaks.
Evidence377 keys per language, four assertions per surface, run on every commit
LimitNative-quality translation is a separate discipline from the engineering that carries it. I build the machinery and the gate; a native speaker still has to read the copy.
-
Tooling for AI agents, where the interesting problem is cost and consent
Gateways that account for the price of a request before it is spent rather than after, queues that stop two jobs from claiming the same GPU, and pipelines that turn video and documents into things an agent can actually use. The hard parts are refusing early and billing correctly when a client disconnects mid-stream.
EvidenceInternal tooling, deliberately not published
LimitThese run for one operator on one machine. They are not a product, they are not multi-tenant, and I would not hand them to a team without a rewrite.
What this page does not show
No team, no on-call rotation, no product with paying users at scale. Everything above was built by one person, and the honest read of that is a range of depth rather than a guarantee of throughput. If you want the version with numbers attached, the GPU work is the one that publishes its losses, and it is the fairest thing on this site to judge me by.