Work

Everything I have built, by the question it answers.

A system that takes instructions from a language model can be wrong at three moments: while the robot moves, before it moves, and in what it was ever possible to ask of it. Every project has its code, its tests and its limits on its own page.

Robotics and software

C++20, Python, ROS2 (Nav2, plugin development), Git and CI/CD, Docker

Simulation

Isaac Sim, MuJoCo, Gazebo, Unity, sim-to-real transfer

AI and perception

PyTorch, TensorFlow, TensorRT, computer vision, LLM integration

Hardware and embedded

JD humanoid, Unitree Go1, Jetson Nano, ESP32 and LoRa, bare-metal C/C++

Try it

Eight of these run here, in your browser.

Six are the project's own algorithm, ported line for line, and recompute the results committed in its repository; one serves the committed certified intervals and grades a score against them; one is a toy with the C++ core's costs. A test in this site's own repository checks each of the seven against its repository on every change.

1 · Safety verifier

Draw a route. The verifier judges it.

Lay waypoints on a real occupancy map, or load what two language models proposed, and the same checks that gate Nav2 say safe or rejected, with the waypoint that failed.

Reproduces all 80 recorded verdicts, two models on six maps

2 · Navigation shield

Watch a bad route get repaired, or halted.

Verify the proposal, plan a recovery on the inflated grid, re-verify with both judges. Seal the goal and it halts rather than invent a path.

Reproduces 50 recorded decisions and 38 recovered paths waypoint for waypoint

3 · ESP32-CAM firmware

Wave at your camera. The firmware decides.

The shipped downsampler and detector, running on your frames in this tab: 192 block means, the cells that changed, the alert line. Nothing leaves the browser.

The exact pipeline from the repository, with a synthetic scene if there is no camera

4 · Plan checker

Edit a robot's plan and watch where it breaks.

Pick a trap, read what a model answered, then change the plan: skip a door, carry the knife into the nursery, and the checker names the step and the rule.

Reproduces all 1,350 recorded verdicts across four models

5 · Tool-call validator

Break a tool call. See which layer catches it.

Rename the tool, break the schema, point at an object that does not exist; the structural and the contract layer answer separately, as recorded.

Reproduces every recorded bucket and all 262 constructed cases

6 · Exact predicates

Move a point by one ulp. Float lies. Exact does not.

Adversarial cases from the committed corpus, with the floating-point and the exact predicate run side by side and the determinants shown.

Every corpus case: exact verdict as committed, float wrong where recorded

7 · Certified intervals

Type a planner's score. The band says what it means.

Thirty-two certified intervals by world and budget: inside the band, below a trajectory that exists, or above a ceiling that nothing can cross.

Served data equals the committed suite row for row, zero violations

8 · D* Lite against A*

Block the robot. Count what each planner redoes.

A wall lands in the route; D* Lite repairs the plan it has, A* starts from nothing, and the cell counts sit side by side. Both checked against Dijkstra.

A toy with the C++ core's exact integer costs; not a repository result

Projects

Start with these four.

The four that say most about how I work. Each has its code, its tests and its limits on its own page.

Catch it while the robot movesllm-nav-shield2026 · LLM plus a verified planner · safe autonomy · Live demo · Film

An LLM proposes a route, a deterministic verifier checks it, and a provably correct planner recovers a safe one. When no safe path exists it halts, in 10/10 such cases, and invents nothing. Code

38/38flawed proposals recovered, none unsafe forwarded
Catch it while the robot movesROS2 LLM safety verifier2025 to 2026 · ROS2 · Nav2 · every proposal replayed in CI · Live demo · Film

A deterministic gate between LLM planners and Nav2. Collision, continuity and map-bounds checks reject hallucinated trajectories before a controller sees them. Code

35/35 · 32/32unsafe plans caught from two models, zero false positives
Catch it before it runsplan-failure-bench2026 · LLM planning · symbolic worlds · Live demo · Film

Every instruction either admits a valid plan or plants exactly one trap, and a deterministic checker scores each model reply with no human or LLM judge. Code

1,350committed model replies, each machine-scored and replayed on this site
Know what was possible at alllegibility-bounds2026 · Motion planning · certified bounds · Live demo · Film

An upper bound on legibility over every admissible trajectory under a path cost budget, not only the ones somebody tried. Code

32/32world and budget pairs certified, zero violations
And seven more
Catch it while the robot movesESP32-CAM motion detector2025 · Deployed · C++, ESP32, LoRa mesh · Live demo

Deterministic firmware for a commercial intrusion-detection prototype, built under contract for Muxtronics. No ML and no vision libraries anywhere. Code

16×12decision grid distilled from 800×600 frames
Catch it while the robot movesROS2 deterministic planning2025 to 2026 · C++20 · Nav2 plugins · CI-verified · Live demo · Film

A* and D* Lite as Nav2 plugins over a ROS-free core, validated against Dijkstra with exact integer arithmetic across 185,237 fuzzed replans. Code

4.3xfaster mean replanning with D* Lite when an obstacle blocks the path
Catch it before it runstoolcall-contract2026 · LLM agents · Python · Live demo

LLM tool calls that parse, type-check and validate, and still break the contract. A schema check, which is what most agent frameworks do, passes 39 of 40 of these calls. Code

32/40pass the contract
Catch it before it runsexact-predicates2026 · C++ · numerical robustness · Live demo

Geometric predicates that cannot be wrong, grown from a real D* Lite bug where two keys that should have been equal differed by the smallest step a double can take. Exactness costs about 2x here, not orders of magnitude. Code

657adversarial cases where float is wrong and exact is right
Know what was possible at alllegible-motion-bench2026 · LLM motion planning · pilot · Film

Legibility against path cost and constraints, computed exactly. Three language models called 116 of 120 trajectories legible, including all 25 that were not physically possible. Code

116/120trajectories called legible by the models that proposed them
Built for peopleSafina Portal2026 · In production · React, TypeScript, Supabase

A school management system that runs Safina International day to day: seven roles, prorated billing, payroll, and an append-only audit trail. I am the sole engineer. Showcase repo

33database tables where every change is logged, with no way around it through the API
Built for peopleHuman-robot companionship study2023 to 2025 · HRI · JD humanoid · MSc thesis

A sim-to-real pipeline on the JD humanoid so a gestural and a spoken condition could be delivered identically to 20 participants; under review at Robotics (MDPI). Write-up (PDF)

95%intent identified correctly from speech, 80% from gesture alone

Ten of these projects are also explained without the jargon, with the same numbers and the same stated limits. A side interest in numerical analysis lives there too: finite-time blowup in one-dimensional fluid models.

All code on GitHubPlain-language guides
Next

Want the reasoning behind these?

hello@munawarkazmi.com