Physics-informed machine learning for geophysical inverse problems
LLM systems constrained by deterministic guarantees
Python · C++ · Rust · TypeScript
PIMSR · Kairos · Closed work · Contact
Neural inversion of magnetotelluric and gravity data, evaluated against the two most-used production MT inversion codes on real USArray stations.
Important
Wins all five Yellowstone profiles against both ModEM NLCG and Occam2DMT v3.0, at roughly four orders of magnitude less compute — milliseconds per profile against 5–210 seconds.
Shift-invariant 2D-forward data misfit, section_nrms_2d, lower is better:
| Profile | ModEM NLCG | Occam2DMT v3.0 | PIMSR U-Net |
|---|---|---|---|
| G | 5.32 | 3.92 | 3.59 |
| H-YS | 5.90 | 4.68 | 4.10 |
| I | 10.98 | 9.26 | 5.62 |
| J | 6.28 | 6.40 | 3.49 |
| K | 6.99 | 6.03 | 4.69 |
| mean | 7.09 | 6.06 | 4.30 |
Setup and reproducibility
Trained on simulated geology, then evaluated on 27 real USArray/EMTF stations in the Yellowstone region (42.5–45.5°N, 108.5–113°W) — no field data in training.
Both baselines were compiled from official sources and driven by scripts inside the benchmark repository, so the comparison is reproducible end to end. Methodology changes and negative results are recorded in the report, including a misfit metric that was retired once it was shown to be an artifact.
Repositories
| Repository | Role |
|---|---|
| pimsr-geogen | stochastic geology model generator |
| pimsr-forward | MT and gravity forward modeling, sensor and noise simulation, dataset builder |
| pimsr-inversion | multi-task neural inversion with uncertainty estimates |
| pimsr-benchmarks | comparison against Occam2DMT, ModEM and SimPEG, plus the full report |
Architecture, specification and implementation of a seven-layer trading system I lead across thirteen repositories in the Kairos-cryptoAI organisation.
The design question is how much authority a language model can be given in a system where mistakes are irreversible.
Important
The answer is enforced structurally rather than by prompt: the model never touches the exchange and never sees a raw number stream. It receives compact pre-validated JSON, every critical action passes deterministic risk filters, and a separate engine executes.
| # | Layer | Control |
|---|---|---|
| 1A | quant-scouts · order book, funding, OI, RSI/MACD | pure math |
| 1B | text-scouts · news/X with local ML pre-filter | LLM, low |
| 2 | router · state machine with hysteresis, picks effort | deterministic |
| 3 | aggregator · fuses quant and sentiment | LLM, medium/high |
| 4 | macro-strategist · allocation, shock-triggered | LLM, extra high |
| 5 | risk-manager · leverage and drawdown limits, breaker | deterministic |
| 6 | execution-engine · atomic orders, EIP-712, CCXT | deterministic |
On failure the system degrades into a local protective mode instead of stopping.
Cost control and supporting work
Cost is a first-class constraint: the router stays cheap by default and escalates to expensive models only when signals conflict, with token accounting in the LLM gateway.
| Repository | Role |
|---|---|
| kairos | umbrella specification and ADRs |
| kairos-core | shared contracts and message bus |
| kairos-persistence | TimescaleDB storage with audit trail |
| kairos-backtest | walk-forward and counterfactual experiment matrices |
| kairos-deploy | deployment and monitoring |
Kept public as a record of a hypothesis that did not survive testing.
structured-latent-hypothesis started from a three-point geometric observation and asked whether the resulting mixed-difference structure could serve as a machine learning principle. The global claim is closed as unsupported: strict affine spacing did not hold, the latent prior did not generalise, and commutator routing lost to simpler support-validation baselines. The work has since narrowed to support-calibrated adaptive routing under context shift.
Other repositories
- Training-Dashboard-Demo — reproducible CIFAR-10 pipeline, config to report, with a baseline/improved ablation
- ProtoSwitch — terminal-first proxy watcher and rotator for Telegram Desktop, in Rust
- asm-atoi-exit-code — x86-64 Linux assembly, written while learning the instruction set directly