I build satellite models from first principles, then try to break them.
ML Architect at Eccoi, working on sovereign AI. Satellite AI Research Trainee at Haizea Analytics, mapping Australia's tree canopy from space. Master of Computing at the Australian National University. Previously 3+ years at Accenture building AIOps tools and explainable AI systems.
I'm a machine learning engineer and architect: from training generative models from scratch to designing AI systems that people can actually understand and trust.
I spent over three years at Accenture building AIOps tooling, explainable AI systems, and internal chatbots. At ANU, I'm going deeper into the foundations: generative modelling, transformers, and human-computer interaction. Now I work on two fronts: as ML Architect at Eccoi, on sovereign AI architecture, and as a Satellite AI Research Trainee at Haizea Analytics, mapping Australia's tree canopy from satellite imagery.
Alongside my studies, I teach Human-Computer Interaction at the ANU as a Casual Sessional Academic, serve as a Student Ambassador for the College of Systems and Society, and represent its postgraduate cohort as the elected ANUSA Postgraduate Representative. I've also published research on AI in material science, electric propulsion, and blockchain; see publications.
Long term, I want to build products that democratise machine learning: from national-scale AI capability to creative solutions for problems in grassroots communities.
// click a skill to find out more
Mapping live tree canopy cover across Australia at 10 m resolution: deep learning models that read four years of quarterly Sentinel-2 imagery and are trained on airborne LiDAR measurements, across more than 13,000 footprints. The goal is accurate amounts of canopy, not just a good-looking map, scored once on a held-out test set against Haizea's existing model. Run as a pre-registered, multi-seed experiment programme where every change has to beat the noise floor, handed over as a clean repository with the model, validation code and a ranked record of experiments.
Architecting the Sovereign AI Community of Practice (SAICOP): defining architectural principles and the system context model, assessing and shortlisting platforms, and proving the conceptual architecture end-to-end, aligned to Australian Government delivery, security, and procurement realities.
Teaching HCI at the ANU from Semester 2, 2026: running tutorials and supporting students through design thinking, prototyping, and evaluation coursework.
Built Ingrain, an AIOps tool generating predictive ML pipelines for automated IT ticket resolution. Integrated Explainable AI and dynamic model-performance reporting, designing the supporting UIs from scratch. Built Quasar++, a ChatGPT API-based chatbot for querying internal documents. Formally Specialized in Data Science & Machine Learning; rated Advanced in Machine Learning, Python, UX Design, Solution Architecture, and Cloud Application Architecture.
Designed and implemented full-stack web systems using .NET, MVC, React, HTML, CSS, and JavaScript; delivered a Hospital Logistics System to manage COVID-19 operational workflows.
Flow matching from first principles: comparing v- vs x-prediction across data dimensions (x-prediction stays stable where v-prediction collapses), plus a MeanFlow implementation with Jacobian-vector-product targets.
x-prediction holds at 32 dimensions where v-prediction breaks.
† designed, implemented and evaluated every parameterisation
A 30M-parameter GPT built from scratch in PyTorch and trained on just 3.7M tokens of five-sentence stories, testing depth vs width and RoPE + RMSNorm (25.77 → 24.67 PPL), plus DPO fine-tuning, in the data-scarce regime.
25.77 → 24.67 test perplexity. Depth beat width; SwiGLU didn't pay off at this scale.
† built, trained and ablated every variant
Stress-testing minGRU's "Were RNNs All We Needed?" claims against a LLaMA-recipe Transformer and a causal gMLP on algorithmic reasoning: a 27-experiment grid showing where each architecture breaks, plus a phase transition in Transformer copy learning.
Transformer ≥ 97% on induction at every length. minGRU: chance (~4%).
† team of four; built the Transformer, designed the experiments
A PWA for sustainable food shopping: real-time barcode scanning, live product lookup via Open Food Facts, allergen/expiry/carbon flagging, and A/B interaction logging for HCI research.
HCI research, demonstrated through a live, deployed app.
A native Android social app (post feed, reactions, DMs, and admin moderation) applying Singleton, Factory, and Iterator patterns across a layered DAO architecture.
† team of five; built the frontend, backend skeleton
A portfolio of human-computer interaction work, including prototyping and AR system evaluation.
An unofficial fork of a macOS menu bar app that tracks Claude usage, adding a second Claude account as a provider of its own across the menu bar, popover, dashboard and Studio. Proposed upstream as pull request #284. For everyday use, get the official TokenEater.
Proposed upstream as pull request #284.
† designed and built multi-account support
A Mac menu bar app that makes a cheap Bluetooth LED bar follow whatever is on screen, the feature Philips Hue sells for about A$550. I reverse-engineered the light's Bluetooth commands, calibrated LED colour against the screen, and shipped it with a website and a release.
The Hue screen-sync feature on a A$49 light.
† reverse-engineered the light, designed and built the app
My first hackathon and first vibe-coded app (ANU, 2025): sales analytics for small businesses. Kept as a benchmark, and revisited in 2026 with tests, a self-scoring revenue forecast, anomaly detection, and fixes for what I first shipped.
Its own demo data: 41% analysed → 100%.
† hackathon team of four; rebuilt it solo
Coursework in deep learning (CNNs, RNNs, transformers, generative models), advanced ML and generative AI (diffusion models, LLMs), statistical machine learning, HCI, and software engineering. Every coursework project on this site was graded High Distinction (HD).
Published peer-reviewed research on electric propulsion for fixed-wing aircraft during undergraduate studies.
Das, M. · DOI 10.30919/esg1183. Also presented in work-in-progress form at RTCMM 2023.
Momaya, V.J., Das, M. et al. · DOI 10.57159/gadl.jcmm.2.5.23071
Karthik, A., Das, M. et al. · DOI 10.30919/es8d573
Elected representative for the CSS postgraduate cohort, advocating for student interests and liaising with faculty and administration.
Representing the College at recruitment events and school visits; supporting prospective-student transition and onboarding.
Trained an ML pipeline on historical patient records for early detection of threatening conditions, reaching ~80% accuracy on malignant breast cancer detection.
Studied sensor-based mineral detection and drafted a feasibility study for automated coal grade determination using computer vision and AI.
Hands-on with just-in-time manufacturing; worked on CNC machine programming and cam design optimisation for higher efficiency using machine learning.
Generative models are my happy place: flow matching, diffusion, and transformers built from scratch in PyTorch. I like knowing what's under the hood, not just calling the API.
Give me a gnarly problem and I'm gone for the afternoon: algorithm puzzles, debugging sessions that turn into detective stories, and systems that finally click at 2am.
Where most of my free time goes: airpower doctrine, procurement politics, and how sovereign capability actually gets built rather than announced. An automobile engineer by first degree, so I read it as an engineering problem first. Hence the Rafale at the foot of the page.
Away from the screen: political history, theology, and an ever-growing queue of long-form analysis. I like understanding why institutions, ideas, and power move the way they do.
Covered institute events and wrote long-form pieces on global issues; op-eds on the Me-Too movement and gun control received recognition and acclaim.
Led teams of five to ten organising cultural-fest events (EQ-IQ, Psychology-101): ideation, logistics, and participant engagement for an inclusive atmosphere.
Regular visits to a home for children with intellectual and developmental disabilities with the Rotary Club; sourced and distributed food, medicine, and rebuilding materials in West Bengal after Cyclone Amphan (2020).
I'm open to opportunities in software engineering and machine learning, based in Canberra, ACT. Reach me by email or on LinkedIn.