Engineering Insights, Benchmarks & Analysis

Technical Notes & Publications

Deep dives on computer architecture, NSCT nationwide standardized testing, LLM GPU serving benchmarks, zero-trust AI security, and systems engineering.

⭐ Featured Deep Dive • Top 1.5% Nationwide

National Assessment & Competency • May 2026 • 6 min read

Scoring 98.5th Percentile in the NSCT: What Fundamentals Actually Matter

Reflections on ranking in the top 1.5% nationwide among 40,700+ final-year CS and engineering candidates across 119 universities. Why industry standardized testing reveals a widening gap in first-principles understanding — from cache hierarchies and bitwise manipulation to algorithmic invariants.

Key Architectural Invariants:

1. Memory Hierarchy Locality:

Understanding spatial/temporal cache line behavior outweighs raw instruction micro-optimizations.

2. Bitwise Invariants:

Masking, shifting, and alignment form the mechanical foundation of all systems code.

3. Algorithmic Bounds:

Rigorous proofs of time complexity and space overhead prevent catastrophic runtime stalls.

#NSCT #ComputerArchitecture #Algorithms #SystemsEngineering

Ahmed Raza

LLM Systems & Inference May 2026 • 8 min

Serving 120B Parameter LLMs on AMD MI300X with vLLM

Benchmarking throughput, concurrency, and time-to-first-token (TTFT) when serving dense open-weight models on unified CDNA3 architecture using vLLM and PagedAttention custom kernels.

Benchmark Takeaway: Unified HBM3 memory allows dense 120B models to reside without tensor-parallel sharding across multiple nodes, reducing inter-GPU communication latency.

#AMD MI300X #vLLM #CDNA3

Hardware Benchmarks →

Cybersecurity & AI Safety Jun 2026 • 7 min

Adversarial Defenses in Autonomous Agent Architectures (CLLMSP)

Key findings from the Certified LLM Security Professional designation: How indirect prompt injections bypass simple regex/filter guards, and why zero-trust capability tokens are mandatory for agent tool execution.

Security Rule: Treat all retrieved text as untrusted user input. Never allow LLM text generation to directly invoke system calls without cryptographically signed authorization tokens.

#CLLMSP #PromptInjection #ZeroTrust

Security Writeup →

Cognitive Science & Software Jan 2026 • 5 min

Building Sovereign Cognitive Tools: Engineering Lessons

Why modern educational and cognitive software needs a "fade-out" design pattern that empowers users to internalize mental models rather than trapping them in gamified notification loops.

Design Thesis: Great tools should aim to make themselves obsolete by building permanent intuition in the user.

#CognitiveDesign #ClientSide #Privacy

Design Philosophy →