What it is
Tolun AI is the first agentic AI platform for mineral exploration. We ingest heterogeneous
geophysical data (gravity, magnetics, IP, MT), run joint probabilistic inversion with full uncertainty
quantification, and deliver drill-ready targets through a chat-first interface a field geologist can operate
without a PhD in computational geophysics.
A coordinator AI orchestrates specialized agents — Inversion InterpreterQC AgentDrill AdvisorReport WriterOrchestrator
— that collapse 20–30 hours/week of geophysicist work into minutes.
Why now
- Mining supercycle + critical minerals: REE, copper, lithium budgets up 40% since 2023; explorers need faster drill decisions.
- Joint inversion matured academically but is un-productized commercially. Geotexera sells it as consulting hours; Mira Geoscience wraps UBC-GIF behind expert-only GUIs. Nobody offers self-serve joint inversion with AI interpretation.
- Agentic AI is reliable enough for domain reasoning today. A "geophysicist-in-a-box" is 2026 feasible, not 2028 bet.
What's live (April 2026)
28
commits shipped in 30 days
540,000+
real data stations processed
5 / 13
agents live / mapped
40–49%
uncertainty reduction (joint vs. single)
6 weeks
from launch to production
- 1D MCMC, 2D Tikhonov, 3D gravity inversion engines (Python + SimPEG-style kernels)
- Bounded joint gravity+magnetics with cross-gradient coupling (Gallardo & Meju 2003)
- DRAM MCMC with adaptive covariance + delayed rejection (Haario 2006)
- Chat interface with geological interpretation
- Live SaaS at api.tolun.ai
Real-data validation
Mountain Pass REE carbonatite (California) — 1,470-station E-W profile, 40×20 mesh, 25 m surface cells,
bounded ±1000 kg/m³, alpha cooled 2⁶→1. Recovered density anomaly consistent with known carbonatite geometry.
Joint gravity + magnetics (Mountain Pass) — cross-gradient coupled, bounded density + susceptibility.
40–49% uncertainty reduction vs. gravity-only.
Bushveld Complex (South Africa) — 410-station W-E profile across 680 km, PGE reef target.
Same bounded algorithm; drill target marked at cross-section maximum.
Technical moat
- Bounded least-squares joint inversion. Unconstrained joint inversion produces unphysical amplitudes (we reproduced this on Bushveld — ±4000 kg/m³). Our augmented stacked system with
scipy.optimize.lsq_linear bounds is a production-grade fix competitors haven't shipped.
- DCT basis parameterization. 240 params → 12 coefficients makes high-resolution MCMC tractable.
- Petrophysical coupling across 22 rock types (density ↔ susceptibility).
- Agent layer — 13 specialized agents with a single coordinator; natively chat-first.
Team
- Fatimah Abdulghafur, PhD — Co-founder / Geophysics. 15 years upstream experience (Maaden); Saudi Vision 2030 mining program. Middle East mining network.
- Alim Polat, PhD — Co-founder / Engineering. Principal LLM Engineer, 10+ years production agentic AI. Just architected SmartDraft at ICON plc (multi-agent platform, RAG over 15,048 docs, 53 days). Previously AI Lead at Handelsbanken — team of 35, fraud F1 0.3 → 0.926, conversational AI 1M+ users. PhD Linköping (Electronics); M.Eng Chalmers (Nanoelectronics); published in Nature and Cell.
Business model
Vertical SaaS. Explorer tier $25K/year · Enterprise $100K/year · Major $250K+/year. No consulting hours.
No services revenue.
Current ask
30-minute technical demo. We'll stream a live inversion on a dataset of your choosing —
Mountain Pass, Bushveld, or your own survey — and walk through how the agent layer interprets it for a geologist.