SkyScout
A Mars scout drone that maps hazards from the air and hands safe routes to a rover.
10.9 mDead-reckoning drift in 14.4 s, before the filter went in
Sankalp Singh
Simulation, estimation, machine learning and AI engineering
I build simulations, then test whether their results hold up.
Illustration of the SkyScout survey, drawn live in your browser from seed 20260808. It is not simulator output.
DemoSkyScout: the scout mission
Demos
The work, running
All demosThe drone flies its survey over the procedural Mars terrain and builds the hazard map the rover plans over.
The drone-swarm simulator running, with the threat classes moving in the airspace the detector watches.
Navigation continuing after the GPS fix is gone, with the estimated track drawn against the truth.
Missions
Flagship work
All 11 projects
Adversarial Swarm Defence Sim
A GAN learns to disguise a drone attack. A detector learns to catch it anyway.
6.0The output scale that fixed a critic winning for the wrong reason
inkless
Markdown to a typeset PDF, with the PDF format written out byte by byte. No dependencies.
0 bytesThe size of requirements.txt, and that is not an oversight
CompliSense
An AI legal-compliance platform for Indian small and medium businesses.
84.8%Macro F1 for BERT named-entity recognition
Research
Two preprints
Read bothA Lead Time Is Not a Detection
Grokking is when a neural network memorises its training data, looks stuck for a long time, and then suddenly starts generalising. Several papers report signals that warn you before it happens, quoting how many steps of lead time they give. This paper audits those claims. Under a rule fixed in advance, most of the reported lead time turns out to depend on how you were allowed to measure, not on the signal seeing anything. One signal survives the strict rule on every seed. A widely cited one fires on none.
- 8/8Seeds where the Fourier signal family led the test-accuracy jump
- 2,550 to 2,650Median steps of lead for the three Fourier signals
- 0/8Seeds where weight norm fired at all
Surrogate Trust Audit
A neural PDE surrogate is a machine-learning model trained to stand in for a slow physics solver. It is much faster, and sometimes it is wrong. Several signals are supposed to tell you when to stop trusting one. I registered a deliberately unintelligent baseline as the primary comparison before running anything: a model of matched capacity that sees only the input, never the physics. It beat every trust signal I tested. That is a negative result, and it is reported as one.
- -0.286PDE residual against the null, 95% interval -0.333 to -0.235
- -0.694Seed-ensemble disagreement against the null
- -0.077Frozen-feature embedding probe against the null