Every public page published on zuwasi.github.io/Public-html-pages, grouped by topic, with the first publication date and the latest update date from git history. Type to search, or click a section to filter.
Dynamic one-page scrolling infographic based on the Sourcegraph CERN case study: animated counters (15M lines of Java, ~300 developers, 18 years of growth), an 18-year timeline, run mode vs upgrade mode change clocks (5-year LHC stability periods vs 1-2 year upgrade windows), four ways code search shrinks technical debt, and the Chris Roderick quote. Built for social sharing.
NEW · MISRA C:2025 · SOURCEGRAPH + PARASOFT APPROVAL DECK
11-slide ESL solution deck for eliminating MISRA C:2025 Rule 7.5 violations across 4 repositories and 400 branches. Opens with the engineering loop as a flowchart: a Parasoft C/C++test baseline scan of one module (~500 findings) extrapolates to ~50,000 project-wide, triggering the decision to automate; Sourcegraph search queries and Batch Change scripts fix every incident; a Parasoft verification scan is the judge, and a non-zero result loops back to script redesign until zero is proven, then scaled to all repos and held by a standing CI gate. Includes Sourcegraph UI recreations, an annotated real Parasoft screenshot, an ESL-style compliance dashboard, and the edge-case safety rails.
Technical review for the C++test product team: implementation, CMake companion workflow, heuristic CUDA rules, CUDA GoogleTest host-coverage evidence, limitations, per-project stress results, and feedback questions.
Customer evaluation deck showing CMake preparation, generated Visual Studio companion analysis, genuine Error List screenshots, CUDA guideline rule scope, result handling, prerequisites, and limitations.
13 שקפים במיתוג ESL: ארבע דרישות של Iroad מהמצגת הראשונה יושמו ואומתו (מקרא למפה, אילוץ קרקע חקלאית, איסור חציית שטח בנוי, והטמעת ספר העקרונות PDDM של FHWA), כל אחת מסומנת "נוסף בעקבות דרישת Iroad". פירוט מקורות הנתונים (Copernicus GLO-30, OpenStreetMap, GovMap) עם תאריכי שליפה, שינויי האלגוריתם לפי התקן (עקומות אנכיות, מדידת חשיפת עצירה, 18 הוכחות Lean 4), סימון נקודות אדם בלולאה על צילומי ההדגמה, והסבר על בקרת גרסאות ושחזור כל גרסה.
ESL slide deck on the completed benchmark: same model (claude-opus-5-5) on both arms, 12 repos, 60 hidden-test-scored bugs, official SWE-sweep Docker verifiers. Amp 4/60 (6.7%) vs Claude Code 2/60 (3.3%), above the official SOTA entry of 4.7%. Introduces SWE-sweep and its leaderboard, quotes the authors' harness ablation, walks through the real diffs the agents got right and wrong, presents the Parasoft C++test control (0-0.3% line-level overlap with ~500k findings), and concludes the harness is the king. Companion pages: linkedin.html (ready-to-post text) and article.html (website article).
12 repositories across 6 languages, 60 real reference bugs, same model (claude-opus-5-5) on both agent arms with equal 1-hour caps and hidden-test Docker verifiers. Final results: Amp 4/60 (6.7%) versus Claude Code 2/60 (3.3%), with cost-tier run 1, per-repo tables, cost model, and provenance.
15 ESL-branded slides reproducing Daniel Lemire's GH Archive workload on a 24-core Core Ultra 9 laptop in isolated WSL. Field projection reached 16.07 GB/s median versus 12.87 GB/s for lazy views at 24 threads: 24.8% higher throughput and 19.9% less time.
Trade-show loop deck in the Amp theme (near-black, cream, orange): the harness is more important than the LLM, then each layer of the AI Coding Agent stack as an animated infographic with its share of the product: User Interface 5%, Agent Orchestrator 15% (planning and control loop), Context & Indexing 15% (repo indexing and code search), Tool Execution 20% (file ops, builds, tests, MCP), LLM Core 15% (just the engine), Verification & Feedback 20%, and Memory & State plus Safety & Guardrails for the remaining 10%. Slides auto-advance in a circle every 6 seconds by default, with a big pause button for talking and an adjustable slide time on the first slide (4 to 12 seconds, steps of 1). Thesis credit: "The Harness Is the Product" (Towards AI).
A workflow proposal prepared for the authors of "Energy Transmission Across Holographic Conformal Interfaces in General Dimensions" (Arav, Bertrand, Chapman, Policastro, Waeber): what the MantiQ Factor process would do with their paper. A 12-slide fit assessment opening with the MantiQ Factor production-line overview, then a six-stage pipeline diagram: freeze the source, normalize claims and conventions, a Wolfram notebook oracle re-deriving the warp-factor transmission formula (Eqs. 15, 29, 154), scoped Lean 4 theorems for the central-charge bounds, a QASM / Classiq circuit artifact executed on a quantum simulator (with an interactive scattering lab), and a public evidence package. Includes a claim-by-claim evidence table and an honest verdict: strong fit for the Wolfram, Lean, and simulator stages; no RTL/FPGA stage applies to a theory paper.
ESL-branded dashboard for the C++ Doctor case repository: all 25 book cases transcribed verbatim and scored line-by-line through Parasoft C++test Professional 2025.2.0 across four configurations, including a custom merged MISRA C++ 2023 + Flow Analysis Aggressive config (277 rules + 147 aggressive flow checkers). Headline: 9 of 25 bugs get a purposeful finding (6 detected outright, 3 flagged on the right line), 12 stay fully invisible, and Recommended Rules alone finds 2. Per-case verdict table, the 13 bug-line findings with verbatim rule messages, rule-mix bars and why the misses need sanitizers. Also includes a full written findings report.
22 interactive slides from two cold Bazel builds of Apollo's cyber and planning modules (24 cores, GCC 7.5), one with GNU ld.bfd and one with mold 2.0.0, with every link action timed through bazel --profile. Whole-build wall time moved +1.1% (462 s vs 467 s), within run-to-run noise. Per link, the median was 16.8% faster with mold and the best single link was 2.94x faster, but links under 1.5 s often got slower: 119 of 124 losses were in that range. Includes all 310 paired link timings, a comparison with the large single-link workloads in Rui Ueyama's arXiv:2608.23228 paper, the setup problems we hit, and when mold is worth adopting.
30 widescreen slides introduce both tools and the test method, show representative planted C defects beside Parasoft's actual rule IDs and messages, and list all 50 valid defects in the adjudication appendices. Parasoft's only scored configuration was Flow Analysis Aggressive, run on the clean and bugged states: 15 strict true positives out of 50 valid defects, plus one partial and one indirect signal (not counted as direct detections). It missed all eight concurrency defects and nine of eleven size-arithmetic defects. SonarQube's DEEP code scan was unavailable — Vortex returned 403 and every requested file was skipped — never zero findings; its local secrets scan completed with zero secrets. Links to the MIT repository, defect catalog, scan protocol, comparison report and detection matrix.
41 ESL-branded slides plus PDF: ESL and CTO Daniel Liezrowice, five project challenges, and one Clang-built application run on the real ThreadX 5.6 Cortex-M4 kernel and on MIT ThreadX 6.5, then built for Cortex-M85. Verified TX5-M4, TX6-M4 and M55 QEMU runs, eleven code-diff discussions, binary-kernel inspection, measured memory charts and histograms. Mobile arrows, Home, a recorded LED/button replay and ESL contact links. M85 execution and real-board performance remain unverified.
Animated customer-demo plan: Apollo's 471,310 C/C++ code lines, six independent PRs, Jenkins, six Parasoft automation servers, optional Incredibuild acceleration, and one DTP. Follow 100 selected findings through Amp-assisted fixes, unit tests, rescanning, and before/after evidence. Proposed targets, not measured results.
By ESL (Engineering Software Lab), a member of Anthropic's Claude Partner Network. What Wazatator is and how it works: an open-source fork of Microsoft Waza that runs your SKILL.md through real Claude Code sessions and scores the result. Code samples for eval suites, tasks, graders and trigger tests, real findings (skills that ask instead of act, skills never invoked, tool-policy violations, MCP usage, cost), a before/after fix, and a replayed live demo.
A humorous lunch-decision app. Choose a city, date, hunger level, hummus style and sources, then watch four fictional experts turn your preferences into a transparent, source-linked verdict. Ten curated places, seasonal scoring and six of Daniel Liezrowice's own posts in the Facebook group “החומוסולוגים”, with a group link and clear evidence limits. Absolutely no actual chickpeas.
Beyond "interrupts vs polling": seven Cortex-M code samples compiled against real STM32F4 and nRF52 headers. Covers the NAPI-style hybrid, DMA with idle-line detection (and its half-transfer gotcha), zero-CPU PPI event routing, the lost-wakeup race in "check, then WFI" with real disassembly, BASEPRI critical sections, interrupt storm guards, and debouncing by polling.
Amadeus Ferro's print_fields() plus a compile-time wire-layout check, built and run on GCC 16.1 through Compiler Explorer. Six compile errors between the P2996 snippet and a running binary with the fix for each, two negative tests, the static-analyzer parse gap, and one MISRA C++:2023 consequence. A reply to a LinkedIn post.
Twelve customer-facing slides: first check available license features with cpptest-license-features, then use the nine-app C++17 test project to verify three scans per container. Includes interactive scheduling and DTP views. Utility compatibility, customer scans and license acceptance remain to be verified in the lab.
20 mobile-friendly slides: distinguish product supply-chain security from CSPM/CNAPP, then compare SBOMator with Black Duck. Includes site-license pricing, ESL’s 60-day integration offer with payment after success, and contact details. Product master-plan capabilities assumed complete.
Eight slides with a live gate-level QPE simulator: exact and approximate phases, eigenstate superpositions, deliberate circuit faults, independent Wolfram vectors, and a scoped Lean resource-count proof. Download reproducible evidence; no hardware or quantum-speedup claim.
Five copyable discussion comments with code examples: counterexamples, observability debt, atomicity, learning loops, and durable code ownership. A six-slide companion to Michal Stawarz's post on AI-assisted engineering.
Six interactive slides: a $5 misrouted ticket versus a $10,000 wrongful account closure, a live expected-loss calculator, copyable Python policy, Jev confidence caveats, and a downloadable infographic. ESL AI SDLC, Daniel Liezrowice.
Seven interactive ESL slides on Jev and cost-aware automation: compare a $20 credit with a $20,000 reimbursement, experiment with expected error loss, copy the Python policy, and download the infographic. Illustrative cases, not measured Jev performance.
Three NXP, three ST, and three AMD/Xilinx-targeting repositories: 18 source-only scans, experimental vendor lifecycle enrichment, downloadable paired SBOMs, and HTTP retrieval trials. Shows unchanged component inventories, coverage gaps, and an obsolete ST board whose MCU still has a longevity commitment.
Daniel Liezrowice, ESL: an interactive walkthrough of context filtering, missing-evidence review gates, and confidently wrong routing. Adjust the threshold to compare errors with deferred work. Synthetic illustration, not measured Jev performance.
Solidity static analysis, finding triage, EVM tests, and Wolfram/Lean 4 model evidence, presented by Daniel Liezrowice of ESL through a 10 test ETH escrow case. Static analysis is core to the proposed workflow; the provider will be an ESL tool or Olympix, subject to readiness and commercial terms. The current Olympix scan remains blocked, with no completed detection claimed. Explore the recorded-evidence tour, workflow image, and supporting SBOMator report (368 components, zero missing licenses, vulnerability findings retained). Not an audit or a production security guarantee.
A critical, evidence-first examination of agent memory claims. Compare stateless, full-context, and governed selective-memory strategies across six interactive scenarios, inspect decision traces, disable safeguards, and download reproducible benchmark results.
The textbook "6 steps of a function call" is a picture of -O0. Real arm-none-eabi-gcc output shows the call vanishing at -O2 (30 instructions become 2), what a forced call looks like (stack alignment padding, -fipa-ra keeping a scratch register live, pop {r3, pc} as the ROP surface), plus an emulator-driven animation of every instruction, register and stack word.
A hands-on companion to Vivek Bhadra's “Beyond Breakpoints”: one reproducible use-after-free defect investigated with AddressSanitizer and GDB, corrected through explicit ownership, and protected by automated regression tests—with a one-click WSL runner and findings dashboard.
A 60-slide evidence-driven journey from the public REED paper through Mathematica, Lean 4, vendor-neutral VHDL simulation, and PYNQ-Z2 implementation—positioned as a reproducible engineering path toward specialized FHE hardware, with physical validation explicitly pending.
MSVC 19.51 /O1 feeds address offsets instead of array values into adc — a value-versus-address optimizer regression. How layered verification (static analysis, unit tests, coverage, compiler diversity, CI gates) with Parasoft C/C++test catches the bad binary before release. Originally published on LinkedIn Pulse.
16-slide technical breakdown mapping every node in a graph-based coding agent state machine to Amp's built-in tools. Side-by-side code, oracle vs static gates, parallel milestones, and a decision framework for regulated industries.
MSVC v19.51 optimizer confuses C++ truncated modulo with Euclidean modulo, miscompiling 6U < (v % 5) for negative inputs. How CERT-C INT10-C, MISRA-C 10.4, and C++test unit tests catch the compiler's own error.
גרסת עורכי הדין: ESL כספקית כלים לתעשיות מפוקחות, לקוחות ולוגואים, SPDF "מי מוודא את המוודא", DataBOM כשכבת הראיות ל-LIA/DPIA — התקנה מקומית air-gapped
One-pager for a free Windows tray utility that copies every saved file's full path to the clipboard, so you can paste the path straight into Amp when you cannot paste the screenshot itself
Pre-scan briefing for an FDA cyber-device submission: directory map, firmware families, HAL and RTOS mapping, 15 Git submodules plus vendored-source inventory, licensing notes, a 10-point scan-module checklist, and safety-critical component flags for the SBOM tooling provider
Animated 13-slide demo with real 1.4.7 screenshots and typed terminal replay: risk dashboard, CI/CD reference verification, and the ENISA SRP drill with 24h and 72h clocks, paste sheet, package manifest and audit chain. Includes an MP4 version (sbomator-cra-reporting-demo.mp4)
Focused 9-page brochure covering complete FDA software-evidence execution: SBOM and CVE remediation, static analysis and bug fixing, unit testing and code coverage, and ALM traceability
Interactive 10-page web brochure: submission-ready SBOM, VEX, HBOM, DataBOM, vulnerability intelligence and postmarket evidence for FDA cyber-device submissions
New in 1.3: OSV.dev malware checks in every scan, real-time endpoint watcher, and the zero-day install-hook heuristic — told through the Aug 2026 keyv/cacheable npm attack
12-slide walkthrough of a closed verification loop: Wolfram symbolic invariant discovery, Lean 4 machine-checked proofs, and differential testing against a compiled eBPF/XDP rate limiter
Complete DevOps guide for running MISRA C 2025 static analysis with Parasoft C++test and the unsupported TI Arm Clang compiler (tiarmclang) using the tiarmgcc_7 compiler family
SBOMator scan report on XFusion: 554 components, CVE matches with EPSS/KEV, VEX triage, and a documented malicious-package false-positive removal (Shai-Hulud alert on conda brotli-python) — interactive HTML with collapsible sections
P2996 print_fields and a compile-time layout check run on GCC 16.1; six compile errors and fixes, negative tests, the static-analyzer parse gap, one MISRA C++:2023 slide
MSVC v19.51 optimizer assumes Euclidean modulo range instead of C++ truncated modulo — and how CERT-C INT10-C, MISRA-C 10.4, and C++test unit tests would have caught the compiler's own error
9-slide driver architecture presentation: 5-module Zephyr design, 8-state device state machine, TX/RX/IRQ data paths, 58 unit tests, and 55% completion status
ניתוח מלא של MISRA C:2023 עם Parasoft C/C++test Professional, כולל V2-fixed: 135 מ-174 הפרות תוקנו, 39 סטיות פורמליות, 48/48 בדיקות יחידה עוברות, דוח תיקון ודוח סטיות
מודרניזציה של מחברות Mathematica לניתוח התפלגות אורכי transcript עם SVD, המעודכנת כעת עם הנתונים הנמדדים שסיפקה ד״ר אורלי אלתר. המצגת משווה בין שלב הסימולציה הקודם לבין הניתוח הנמדד, מציגה את שינוי אורכי ה-transcript ב-GBM, ומבחינה בין תצפיות אמפיריות, אימות נומרי ב-Wolfram והוכחות פורמליות ב-Lean 4.
Full SBOMator scan report on WebGoat: 203 components, dependency graph, CVE matches with EPSS/KEV, lifecycle status, and remediation recommendations — interactive HTML with collapsible sections
SBOMator scan report on XFusion: 554 components, CVE matches with EPSS/KEV, VEX triage, and a documented malicious-package false-positive removal (Shai-Hulud alert on conda brotli-python) — interactive HTML with collapsible sections
No pages found. Try a different term or clear the search.
Engineering Services
C/C++, embedded systems, MISRA compliance, software supply-chain security, AI assessment, formal verification, research-to-hardware prototyping, and the ESL AI LogicLab product portfolio, each supported by public engineering evidence.
