Skip to content
View DrMkdaddy's full-sized avatar
3 followers · 15 following

Achievements

Achievement: QuickdrawAchievement: YOLOAchievement: Pull SharkAchievement: Arctic Code Vault Contributor

Achievements

Achievement: QuickdrawAchievement: YOLOAchievement: Pull SharkAchievement: Arctic Code Vault Contributor

Block or report DrMkdaddy

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
DrMkdaddy/README.md

Noor Mkdad

Platform and systems engineer. I build Linux infrastructure and LLM serving systems.

Email · Resume · stanzaapi.com · Engineering notes

I'm looking for a full-time platform, backend, or systems role. I'm also interested in AI infrastructure work.

Selected work

Project What it is Result
Local inference stack OpenAI-compatible gateway for one RX 9070 XT, with cost-aware routing, MCP, Prometheus, and a Nix deploy. $0.19–0.41 per million tokens; 424 tok/s at 8 slots.
KV-serving scheduler Go scheduler for llama.cpp with a hard KV budget: block allocation, admission control, and preemption. 120% of capacity; high-priority TTFT p99 74 ms under mixed load.
Tiered KV cache Valkey L2 for an LLM KV cache, with content-addressed blocks and LRU spill. Fixed pool served 4/200 sequences; tiered served 200/200. Restore beats prefill above ~4k tokens.
llama.cpp Vulkan kernel K-quant dequantization hoisting and 8-row blocking in mul_mat_vecq for RDNA4. +13.4% aggregate at 8 slots; eight concurrent greedy outputs bit-identical. PR #31022.
X12 parser benchmark Public benchmark for the zero-regex X12 streaming parser behind Stanza API, compared against node-x12 and x12-parser. 255,000 records in 14.2 s inside a 25 MB V8 old-space cap.
Platform infra OpenTofu and NixOS configuration for a Proxmox lab, with recovery runbooks. Provisioning and host configuration are committed; recovery is documented per failure mode.

Stanza API

I built and operate Stanza API: 12 production APIs for healthcare claims, payments, and supply-chain data. It has SDKs in four languages and an MCP server.

Tools

Go, TypeScript, Python. NixOS, OpenTofu, Proxmox, Cloudflare Workers. llama.cpp, Vulkan, AMD RDNA4.

Other repositories

Qwen MoE streaming · Speculative decoding · Quantization quality · Paged attention

Upstream

Merged: stable-diffusion.cpp #2293, rerun #12943. Open: llama.cpp #31022.

Pinned Loading

  1. local-inference-stack local-inference-stack Public

    Self-hosted LLM inference serving stack on AMD RX 9070 XT: OpenAI-compatible API, MCP server, benchmarks

    JavaScript

  2. kv-serving-scheduler kv-serving-scheduler Public

    Bounded-KV serving scheduler for llama.cpp: block allocation, admission control, preemption, and measured 120%-capacity behavior on one 16 GB GPU.

    Go

  3. valkey-kv-l2 valkey-kv-l2 Public

    Valkey L2 for an LLM KV cache: 200 of 200 sequences served against 4 with a fixed pool; restore beats prefill only above ~4k tokens.

    Python

  4. llama-vulkan-matvec-rdna4 llama-vulkan-matvec-rdna4 Public

    llama.cpp Vulkan matvec optimization for RDNA4: K-quant dequant hoisting + 8-row blocking. +13.4% aggregate at 8 concurrent slots; single-stream unchanged.

    JavaScript

  5. platform-infra platform-infra Public

    Declarative infrastructure for a private Proxmox VE lab: OpenTofu provisioning, NixOS flake configuration, and documented failure runbooks.

    Nix

  6. StanzaAPI/benchmark-harness StanzaAPI/benchmark-harness Public

    500 MB of X12 claims in 14.2 s inside a 25 MB V8 cap, benchmarked against node-x12 and x12-parser.

    TypeScript