Skip to content
View munawarkazmi's full-sized avatar

Block or report munawarkazmi

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. 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
munawarkazmi/README.md

Munawar Kazmi

Robotics & AI Research Engineer

I build robots that navigate safely, AI that runs on the edge, and software that real institutions depend on every day.

MSc AI & Robotics (Commendation) · University of Hertfordshire

Portfolio Preprint


4.3x 0 / 2,151 32 / 32
faster replanning
D* Lite vs A*, 200 seeded trials
unsafe trajectories missed
by my LLM safety verifier
certified legibility bounds holding
every world, every budget

My work sits where autonomy meets responsibility: human-robot interaction, sim-to-real transfer, and keeping learned systems honest before they touch hardware.

Every number below is checkable. Each one is produced by committed code and re-verified in CI on every push — clone the repo, run one command, get the same number. Where a result has not survived review, the retraction is published next to the claim it replaced.

Featured work

Project What it is
legibility-bounds Certified two-sided bounds on how legible a robot trajectory can be under a path budget — quantified over every admissible trajectory, not the ones somebody searched. 32 world-ceiling pairs, 0 violations. RA-L draft in paper/.
plan-failure-bench How LLM planners fail at robot tasks, not just how often: 60 trap-labelled instructions, ground truth decidable end to end, no human or model judging anywhere. 548 CI-enforced tests. Citable preprint: DOI 10.5281/zenodo.21756817.
llm-nav-shield An LLM proposes a trajectory, my verifier checks it, my planner recovers a safe one, and the system halts when none exists. Replaying 40 committed proposals: 38/38 flawed plans recovered, 0 unsafe forwarded.
ros2-llm-safety-verifier Deterministic safety gate between LLM planners and Nav2. Over 40 real scenarios it caught 35/35 unsafe qwen2.5-7B and 32/32 unsafe llama-3.3-70B trajectories — zero misses, zero false positives, microsecond latency.
ros2-dynamic-path-planning A* and D* Lite as Nav2 plugins over a ROS-free C++20 core. 4.3x faster replans on average (11x median) in a fair seeded benchmark, validated against Dijkstra ground truth on 185k fuzzed replans.
legible-motion-bench What clarity costs a robot in safety, measured exactly with no human or model judging. Two models called all 80 of their trajectories legible; 25 of those were not physically possible.
toolcall-contract Two-layer validator for LLM tool calls. The structural layer agrees with pinned jsonschema across 25,000 differential cases; the semantic layer catches what it cannot — 7 contract breaks where the schema layer saw 0.
exact-predicates Geometric predicates that cannot be wrong, grown from a real D* Lite key-tie bug. 657 committed adversarial cases where CI asserts the float version is wrong and the exact one is right. Exactness costs ~2x, measured.
degregorio-blowup Finite-time blowup in a 1D model for 3D Euler. Regressing an unknown constant against a known one cancels the discretisation error and collapses the spread 800-fold: beta = 3.0024227 +/- 5e-6, excluding the natural guess of exactly 3.
esp32-cam-motion-detector Deterministic motion-detection firmware from a commercial prototype (Muxtronics 2025). No ML, no vision libraries, CI-compiled for the target board on every push.
safina-portal-showcase School management system in production for a real institute: seven user roles, prorated billing, payroll, and an append-only, trigger-written audit trail. Built, shipped, and operated solo.

Toolbox

Robotics and systems

C++ Python ROS2 Docker Git

Edge AI and embedded

TensorRT Jetson ESP32 LoRa

Full-stack

React TypeScript PostgreSQL Supabase

Currently

  • Extending plan-failure-bench's k=5 sampling protocol to the remaining grid cells — the single change that would most strengthen its claims
  • Preparing the legibility-bounds RA-L draft for submission
  • Running and evolving a production school platform serving students, teachers, and staff daily
  • Open to research collaborations and PhD opportunities in robotics and trustworthy AI

Pinned Loading

  1. plan-failure-bench plan-failure-bench Public

    Benchmark measuring how LLM planners fail at robot tasks, not just how often: trap-labelled instructions with machine-checked ground truth, differential-tested against pyperplan, no human or LLM ju…

    Python 4 1

  2. legibility-bounds legibility-bounds Public

    Certified two-sided bounds on how legible a robot trajectory can be under a path budget: the upper end rules out every admissible trajectory, not just the ones somebody searched, and the lower end …

    TeX 1

  3. ros2-dynamic-path-planning ros2-dynamic-path-planning Public

    A* and D* Lite Nav2 planners over a ROS-free C++20 core, validated against Dijkstra ground truth with exact integer arithmetic. Seeded 200-trial benchmark: D* Lite replans 4.3x faster on average (1…

    C++ 9 1

  4. llm-nav-shield llm-nav-shield Public

    Neurosymbolic navigation shield: verify LLM trajectory proposals, recover with a provably-safe classical path, or halt when none exists. Composes two verified cores as pinned submodules; qwen2.5-7B…

    C++ 2 1

  5. ros2-llm-safety-verifier ros2-llm-safety-verifier Public

    Deterministic safety gate between LLM planners and Nav2: footprint, continuity, and map-bounds checks before anything reaches a controller. Caught 35/35 unsafe qwen2.5-7B trajectories, zero misses,…

    C++ 7 2

  6. safina-portal-showcase safina-portal-showcase Public

    Production school management system running a real institute every day: seven tailored roles, admissions, attendance, homework, cash-based billing with proration, payroll, library, facilities, digi…

    TypeScript 4