Skip to content
View moclamzw's full-sized avatar
1 follower · 1 following
  • Dubai, UAE

Block or report moclamzw

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.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. 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
moclamzw/README.md

Malcom Mudhungwaza

Senior AI Engineer — GenAI, LLM and agentic systems PhD candidate — AI safety & alignment · Dubai / Abu Dhabi, UAE

I build LLM systems and then measure whether they actually work. Most of what follows is the second half of that sentence.

Currently researching whether safety guardrails survive translation across language and modality — and whether the benchmarks we use to check are measuring refusal or just reading comprehension.

What I'm working on

A set of repositories, each built to answer one measured question rather than to demonstrate a technology. Every number in them is generated by a committed script, and CI fails if a result drifts from the code that produced it.

Inference & serving

The question
attention-kv-cache-from-scratch What did grouped-query attention actually buy, in concurrent sequences?
llm-serving-benchmark Which serving config meets a stated SLO at the lowest cost per million tokens?
prefill-decode-roofline Where does measured bandwidth fall short of theoretical peak, and why?
quantization-accuracy-curves How much does the calibration set choice change measured degradation?
speculative-decoding-when-it-pays At what batch size does speculative decoding stop paying for itself?
moe-vs-dense-serving-profile Why do active parameters mislead capacity planning for a sparse model?
context-extension-and-long-context-eval How far short of advertised context does effective context fall?

Evaluation & safety

The question
llm-eval-harness How far can an LLM judge be trusted, measured against human labels?
guardrail-transfer-study How much of the apparent cross-modal safety gap is a reading-ability artifact rather than an alignment gap?
rag-eval-retrieval-vs-generation Would a single end-to-end RAG score have hidden a real regression?
agent-eval-trajectory-vs-outcome How often does an agent reach the right answer through a wrong process?

Post-training

The question
preference-optimization-landscape Which preference-optimisation method for which situation, and what does each give up?
rl-environment-for-rubric-graded-tasks What reward hack did the agent find, and what environment change closed it?
sft-loss-masking-and-packing How much does packing without attention-mask correction actually cost?
peft-lora-memory-and-quality At what rank does quality plateau, and does the memory math match reality?
synthetic-data-pipeline What fraction of a generated set is contaminated against the eval set?

Agents, retrieval & systems

The question
mcp-server-and-agent What is each agent topology's failure rate, and does supervisor actually beat single-agent?
retrieval-internals-and-tuning At what latency budget does re-ranking stop being worth it?
doc-extraction-ocr-vs-vlm Which extraction approach per document class — and does constrained decoding beat a retry loop?
llm-client-kit Where does the wall-clock actually go when you fan out LLM calls?
inference-gateway At what monthly request volume does self-hosting beat the API?
voice-loop-latency-budget Which hop dominates perceived voice latency?

Bold = flagship. Several are early — each README states plainly what is built and what is not, because a status table beats a guess.

Background

  • Now — Senior AI engineer on healthcare AI: multimodal document pipelines, agentic orchestration, human-in-the-loop safety layers for automated clinical-financial decisioning, self-hosted inference on vLLM.
  • PhD (in progress) — Attenuation of Safety Alignment Across Language and Modality in Multimodal LLMs. Two preprints in preparation: MCP multi-agent orchestration, and cross-lingual/cross-modal guardrail transfer.
  • Before — travel and booking platforms at scale (React/TypeScript, FastAPI, C++ microservices); autonomous-vehicle perception (YOLO, ROS2, sensor fusion, SLAM, TensorRT edge inference).
  • MSc Computer Science (distinction) · BTech Robotics & Automation

Stack

Core Python · TypeScript · C++ · SQL LLM LangGraph · MCP · vLLM · Hugging Face (Transformers, TRL, PEFT) · PyTorch Data PostgreSQL/pgvector · Redis · FAISS Infra Docker · Kubernetes · FastAPI · OpenTelemetry · GitHub Actions

Elsewhere

This account dates from July 2026, after my previous GitHub account was compromised. Work here begins from that rebuild.

Popular repositories Loading

  1. attention-kv-cache-from-scratch attention-kv-cache-from-scratch Public

    Attention, MHA/MQA/GQA and the KV cache implemented from first principles, with the memory arithmetic that explains why GQA exists

    Python

  2. llm-eval-harness llm-eval-harness Public

    An LLM-as-judge harness whose headline result is how much you should distrust it.

    Python

  3. llm-client-kit llm-client-kit Public

    The provider-agnostic async LLM client the rest of the portfolio is built on: bounded concurrency, retries with jitter, deadline propagation, cost accounting, record/replay cassettes.

    Python

  4. rag-eval-retrieval-vs-generation rag-eval-retrieval-vs-generation Public

    A RAG evaluation that measures retrieval and generation separately, and finds a case where generation faithfulness improves while retrieval recall gets worse.

    Python

  5. prefill-decode-roofline prefill-decode-roofline Public

    The physics before the engines: measured arithmetic intensity, prefill compute-bound vs decode memory-bandwidth-bound, with the roofline plot.

    Python

  6. sft-loss-masking-and-packing sft-loss-masking-and-packing Public

    One real SFT run where the interesting result is the ablation: prompt-token loss masking on/off, and sequence packing with a correct vs a naively-wrong attention mask.

    Python