I build the measurement tools I kept needing and couldn't find.
muneebs.space · ORCID · LinkedIn · @turing
Four packages I maintain end to end — tests, docs, CI, and a release I can point at.
| Package | What it does | Release |
|---|---|---|
| datasets‑lint | ruff for training data — duplicates, PII, eval-set contamination, schema drift, label leakage, with --fix codemods and SARIF output |
  |
| naqsha‑cli | Architecture diagrams as code — lint, JSON schema, static export, and a self-contained viewer with zoom and copy-PNG |   |
| refilm | Visual regression for DOM state — SVG+PNG frames, a watch loop for development, and accept to promote baselines with an audit trail |
  |
| relabeld | Deterministic re-labelling for annotated screenshots — recipes, diff reports, screenshot evidence in CI |   |
On GitHub Marketplace — the same tools, running where the code already is:
4 packages · 3 Marketplace Actions · 500+ tests · 92% coverage on datasets-lint
muneeb@horizon ~ $ cat focus.txt
> LLM security prompt injection, quantisation-induced attack surface
> Evaluation benchmarks that report what aggregate accuracy hides
> Quantisation what survives compression, and what silently flips
> Product engineering Chrome extensions, MCP servers, self-hosted tools| Paper | What it measures | |
|---|---|---|
| Quantized Guardians — Accuracy-Neutral Decision Churn in Compressed Prompt-Injection Detectors | F1/AUROC hold steady while which inputs get blocked flips | DOI |
| Same Accuracy, Different Verdicts — Quantization-Induced Verdict Churn in Small LLM Smart-Contract Auditors | 175 Solidity contracts, F16 → Q8_0 → Q4_K_M, up to 45.7% verdict flips | DOI |
| SENTINEL-LLM — Practical Multi-Layer Defense for Adversarial Prompts | 7 PoC prompts, 32 real attacks (JailbreakBench / HarmBench / OWASP) | preprint + code |
| TokenAttackKit — Unicode-Level Tokenization Attack Detection | 11 attack families, 0.98 F1 heuristic detector, CPU-only | code |
| Project | Stack | Description |
|---|---|---|
| localbench | Python FastAPI |
Benchmark console for local LLM servers — TTFT, prefill/decode tok/s across 1K–64K contexts, concurrency sweeps, GPU telemetry |
| pin-saver | JavaScript MV3 |
Download a whole Pinterest board as ZIP or PDF — zero dependencies, live on the Chrome Web Store |
| linkedin-mcp-pro | Python MCP Playwright |
MCP server exposing 54 LinkedIn tools to any agent runtime |
| llm-quantization-benchmarks | Python |
Accuracy delta, per-input decision churn and per-layer sensitivity across FP32/BF16/INT8/INT4 |
| agentic-rag-pipeline | Python |
Hybrid BM25 + dense retrieval, reciprocal rank fusion, query rewriting, grounding checks |
| tinyllm-eval-suite | Python |
Reproducible eval harness for small models — task suites, scoring, regression tracking |
Engineering breadth — 10 more repos
speech-to-text-pipeline time-series-anomaly fraud-detection-system
neural-architecture-search mlops-experiment-tracking synthetic-data-generator
multimodal-vlm-training medical-imaging-diagnosis autonomous-rl-agent
edge-vision-tensorrt
Practical AI Security Series — Kindle · Paperback
01 Quantum Computing with AI · 02 LLM Security and Prompt Injection
03 AI and Blockchain · 04 AI and Cyber Security
muneeb@horizon ~ $ cat stack.toml
[core]
python = "primary — PyTorch, FastAPI, numpy"
js = "TypeScript, Chrome MV3, React"
infra = "Linux, systemd, Nginx, Docker, Cloudflare"
[ai]
models = "transformers, llama.cpp, vLLM, quantization tooling"
agents = "MCP, tool-calling, RAG, evaluation harnesses"Writing code, then measuring what it actually does.
muneebs.space · GitHub · ORCID · LinkedIn