Scientific software · AI evaluation · Applied research
From physical models to reliable AI.
Md Ashraful Islam / Ash
I turn scientific questions into testable prototypes, then make their assumptions, reference checks and limitations visible. I’m building toward empirical AI evaluation from a foundation in applied science.
Three working educational demonstrations. Each has a defined model, a reproducible check and a clear boundary on what the result establishes.
01
Numerical verification / Working demonstration
Heat Solver Lab
How can a numerical approximation show us where it is wrong?
An explicit finite-difference model of one-dimensional heat diffusion, compared with a known analytical solution. The interface exposes boundary conditions, stability and grid-point error.
My contribution: problem framing, AI-assisted implementation direction, validation planning, output review and scientific explanation.
At t = 0.2, halving grid spacing reduces maximum absolute error by approximately four times.
α = 0.4, r = 0.4; nondimensional values
Grid intervals
Max. absolute error
20
1.0358 × 10⁻³
40
2.5823 × 10⁻⁴
80
6.4514 × 10⁻⁵
This checks one smooth problem against its exact solution. The current implementation is a classical solver; a trained PINN and physical calibration remain outside its scope.
02
Assumptions & failure cases / Synthetic model
Conveyor Sensing Bench
When does correcting a systematic error make one prediction worse?
A deterministic model separates delayed observation from future-target position. Controls expose constant speed, occlusion, latency and imposed disturbance.
My contribution: model scoping, AI-assisted implementation direction, comparison design and communication of the assumptions.
With speed 40 mm/s, zero latency and occlusion, and a 12 mm disturbance amplitude, the model contains a time point where absolute error rises from 4 mm to 12 mm after correction.
The correction removes systematic offset; the remaining instantaneous disturbance can still increase total error.
Known geometry and timing are supplied to the model. This is synthetic motion, with no camera inference, trained vision, robot control or measured sensor-accuracy claim.
03
Search & verification / Interactive teaching model
Boolean Verification Lab
What is the difference between finding an answer and checking one?
A small Boolean-satisfiability lab shows assignments, clause checks and literal-inspection counts. Reviewers can step through a search, change the preset and export a run as JSON.
My contribution: concept definition, AI-assisted implementation direction, interface review and explanations of complexity limits.
A structured six-variable example takes the displayed enumeration through 64 candidates and 384 literal inspections. Reading its unit clauses reveals an immediate shortcut.
The example illustrates why one algorithm’s search cost cannot establish a lower bound for every algorithm.
Finite teaching examples and operation counts. The lab makes no claim to resolve P versus NP or measure general SAT difficulty.
All three model files are public. The reproduction notes give the settings, commands and limitations behind the checks above.
Independent checks for AI-generated physical simulations.
When do explicit physical constraints and independent reference checks help a verifier detect mistakes in AI-generated scientific work?
The current demonstrations provide bounded starting examples. The AI evaluation study is proposed and would be scoped with research collaborators.
Define checkable tasks.Start with heat diffusion and independently specified reference solutions. Freeze task definitions and separate development tasks from held-out cases.
Compare verification approaches.Compare direct generation, structured assumption recording and independent verification using fixed model versions and comparable budgets.
Introduce controlled mistakes.Vary units, initial conditions and boundary conditions. Measure correctness, false acceptance and failure detection, with uncertainty reported.
Publish the full record.Retain task definitions, prompts, settings, analysis and unsuccessful cases. Test whether the findings survive beyond the initial task family.
A bounded, reproducible study to be scoped with collaborators, with task definitions and evaluation criteria agreed before data collection.
Applied-science graduate and founder of MAI Physics, developing interactive scientific software with AI-assisted implementation. My work connects explicit assumptions, reference solutions and inspectable results. I am seeking a transition into empirical AI research, with a focus on evaluating scientific reasoning, understanding failure cases and making verification useful to people.
MAI Physics
Founder and independent project lead · 2026–present
Develop a platform for interactive experiments, reproducible models and scientific storytelling. Define project questions, direct AI-assisted implementation, review numerical comparisons and explain the assumptions and limitations of each demonstration.
Published a one-dimensional heat-diffusion lab that compares a finite-difference solution with an analytical reference and displays numerical error and stability information.
Developed a synthetic conveyor-sensing demonstration in which known geometry and motion make latency, prediction and disturbance errors inspectable.
Created a Boolean satisfiability visual lab that separates candidate search from verification, with step controls, operation counts and JSON export.
Connect mathematical models, user controls, evidence and explanations so readers can examine how a result was produced.
Hansal Group of Companies
Technology business development and project liaison · 2024 onward
Contribute technology proposals, project planning and research communication.
Supported 2025 RMIT work-integrated-learning coordination, technical scope and documentation for agritech data-collection and smart-irrigation concepts. Read the project note →
MetaCog Build Pty Ltd
Technical project contribution · 2026
Contributed requirements, workflow definition and project documentation for AI-enabled construction and project-management concepts.
7-Eleven
Retail assistant / team member · Current
Customer-facing retail work alongside independent technical development, with practical communication and day-to-day operational responsibility.
04 / Take the work with you
The portfolio reading pack.
Readable PDFs for offline review, with public project links and a clear account of the evidence.
This portfolio and its presentation were prepared with AI assistance from existing career materials and project artifacts. The project descriptions identify the implementation scope, Ash’s contribution and the limits of the evidence.
My role spans question selection, requirements, AI-assisted implementation direction, validation planning, output review and scientific communication. The proposed study is presented separately from completed demonstrations.
Sources & scope
Project evidence: the linked public model files and reproducible calculations dated 5 November 2026.
Career summary: existing candidate career materials, updated for this portfolio.
For research teams, engineering teams and collaborators seeking clear, inspectable technical work.
Continue the conversation
Questions are welcome.
Ask about a model, a check, a limitation or the proposed study.
