Software & ML engineer · Geoscientist · Kraków

Geoscience, machine learning and AI for data that is hard to use.

Plate 1 · Synthetic seismic section · Automatic reflector tracking
Tracked reflector Move over the section to track · Click to keep a pick

Try it. A working reflector tracker, running in your browser. It follows the reflector under your pointer and stops where the signal is too weak to trust.

Months → < 1 min

Boulder picking on sonar, with a network I trained.

200+

Seismic volumes handed over in days, for Shell.

1,100 km²

North Sea study. The licence was awarded.

3 weeks

To a first working interpretation application, solo.

Selected work

The figures are schematics drawn in code, and they respond to your pointer. Client data stays with the clients.

Plate 2 · Side-scan sonar · Schematic
Hand and model Model only

01Machine learning · Offshore wind

A detector that found boulders the manual picks had missed

Offshore wind sites must be cleared of boulders, and picking them by hand on sonar takes months. I trained a convolutional network on about 30,000 hand picks. It found boulders those picks had missed.

< 1 min

On one GPU, against months by hand.

~30,000

Hand-picked boulders for training.

Plate 3 · Handover manifest · Schematic

02Data engineering · Shell · NAM

A divestment's worth of seismic data, delivered in days

When NAM's offshore assets were sold to Tenaz Energy, Shell's seismic had to leave an internal format and arrive as SEG-Y. My pipeline ran unattended and delivered it in days. A later version, for velocity models, is run by an AI agent.

200+

Volumes, more than 50 TB.

½ day → < 1 h

Per velocity model.

Plate 4 · Linked map and section · Schematic
Tracked horizon and point picks

03Software · Built solo with AI agents

An interpretation application, built solo and in daily use

Interpreters at my employer needed software built around their own workflow. I designed and built it with AI coding agents: linked map and section, automatic tracking, boreholes, GIS, surfaces.

3 weeks

To a first working version. In production since.

Plate 5 · Structure map · Schematic, not the client's data
Prospect Lead

04Interpretation · North Sea

An exploration licence, awarded

I led the interpretation and was first author of a petroleum-system study in the Norwegian–Danish Basin for PGNiG Upstream Norway. It fed an APA licence application, and the licence was awarded.

1,100 km²

Interpreted.

3 + 4

Prospects and leads.

Plate 6 · Fitting a cut into a rough stone · Schematic
Accepted cut Inclusion

053D reconstruction · Manufacturing

From 360° images of a rough diamond to a cutting plan

No geoscience here. Software for a diamond company in India that rebuilds a rough stone in 3D from tomographic images, maps its inclusions and plans the cuts. It reached a working prototype.

360° → 3D

Images to cutting plan.

Solo

Built mostly by me.

Plate 7 · Airborne LiDAR profile · Schematic
Stem axis and diameter at 1.3 m

06Point clouds · Forestry · Roads

Trees and roads, measured from a point cloud

Two toolsets on airborne LiDAR: trunk diameter and straightness for forest inventory, and deformation in road surfaces. Research work at Grupa GeoFusion.

2 toolsets

Forestry and roads.

Airborne

From aerial scans alone.

Seismic hazard for a nuclear plant site

Three years. My scripts built the fault catalogue.

Report figures and 3D scenes, generated

Every interpreted profile as a client-ready figure, on about 15 projects.

Velocity models for depth imaging

For PGNiG and LOTOS, with RTM, VTI and TTI variants.

Marine survey processing for offshore wind

Sub-bottom seismic for RWE site surveys, with my own tools.

Format forensics

Seismic recovered from undocumented binary formats.

Method

How one person ships this much

  1. AI agents are my default tools. I design, they type, I review and test.
  2. I read the data first. Ten years of interpretation tells me when a clean result is an artefact.
  3. The work is handed over. Tested, documented, and run by your team without me.

What I take on

I work at WIDMO Spectral Technologies, and take on a few scoped projects a year, each with a defined deliverable.

A geoscience project, delivered

Interpretation, modelling and data management, start to finish.

AI inside your company

RAG over your own documents, and agents built into a real workflow.

A model or tool your team keeps

Detection on images, volumes and point clouds; converters and pipelines, tested and handed over.

A prototype that settles it

A short build on your real data, before you commit a budget.

The same method, on my own records

I keep four knowledge bases for myself: plain files, written rules, and an agent that does the work and shows its evidence. They are shown here on demo data.

Plate 8 · Work knowledge base
Network graph of invented people and organisations, coloured by role
---
name: Signe Aasen
role: Principal geologist
org: halden-geoscience
relationship: client
flags: [client, satisfied-with-delivery, referral-route]
first-contact: 2021-06
---

Commissioned the structural study; gave a written reference.
Knows Marte Halvorsen.

# The graph is built from records like this one. A person is linked to an
# organisation only where the record says so, never by guessing from the notes.
Demo data · names and values are invented · the tool and the structure are the ones I use

01Contracts · Projects · People

One set of files behind the CV, the profile and this page

Projects, people and organisations are plain files with a fixed schema. An agent keeps them consistent and writes the CV, the LinkedIn text and this page from them. The graph is one of its tools.

45

Projects on record, each with its numbers.

Plate 9 · Health knowledge base
A year of invented training, weight and health events on one timeline Map of the data stores, their row counts and how they connect
Demo data · a fictional person, every value invented · the tool and the structure are the ones I use

02Wearables · Scale · Food log

An agent that may not state a value it has not read

Data from a watch, a scale and a food-logging app I wrote lands in files with a fixed schema. The agent syncs it and writes check-ins that cite the file and date behind every claim.

19

Connector modules, across four data sources.

18

Modules with a built-in self-test.

Plate 10 · Workstation knowledge base
Photo of the workstation with the graphics card selected and its specification and measurements listed Six recommendations that measurement overturned, each kept in place with the file that records it
The hardware, the measurements and the tool are real · prices, invoices and live machine state are invented

03Hardware · Tuning · Maintenance

A workstation with a measurement behind every setting

I built the machine this work runs on. An agent maintains it under written rules: measure before changing anything, and never record an unknown as a pass. Its weekly check has caught a stale backup and a stopped fan service.

38

Decisions recorded. Seven were retracted when a measurement overturned them.

9 weeks

Of unattended weekly checks, each verified from its log.

Plate 11 · Petroleum exploration knowledge base
The library's command line looking up, citing and searching facts, then rejecting a quote in which one digit was changed Network of sources and facts with one fact open: its citation, its quote and an unreviewed flag Five invented sources on one vitrinite reflectance axis, disagreeing on where the oil window starts and ends
Demo data · invented sources and facts · the tool and the structure are the ones I use

04Sources · Facts · Citations

A library an agent has to quote, not remember

Books and papers on exploration methods are stored page by page, and each threshold or equation is a fact with a quote from its page. A program rejects any quote that is not on the page it cites, so an agent answering from it cites rather than recalls.

1,247

Facts in 38 sources. Every quote matches its page, with no failures.

680+

Facts re-read blind from the page image by a second agent.

Record

Work delivered for or to Shell, NAM, Tenaz Energy, TNO, PGNiG Upstream Norway, PGNiG, LOTOS and RWE.

Published

Worked at

  • WIDMO Spectral Technologies, Data Architect
  • Shell, Process Data Engineer
  • GEO-ENERGY INNOVATIONS, seismic interpretation
  • Grupa GeoFusion, R&D, geophysics and IT
  • AGH University, senior technical specialist

Studied

  • MSc, Petroleum Geology, AGH University.
  • Engineer's degree, Mining and Geology, AGH University.
  • Doctoral studies, AGH. Coursework completed.

Tell me what the data is and what you need out of it.

piotr.hadro@gmail.com

Or find me on LinkedIn.

Piotr Hadro
Kraków · English or Polish