IanMatejka

1. CLARISSA, a closed harness

   Writes, runs and analyzes reservoir simulations from natural language on the open-source OPM Flow engine. When a run fails, it diagnoses the deck, revises it and resubmits.

2. Fences, not rails

   The model is not marched through a fixed workflow. It reasons freely inside fences set by physics; when it strays, a deterministic guard fires (PVT consistency, relative-permeability endpoints, input checks) and the agent corrects itself.

3. PetroScript, an open ecosystem

   A typed Python DSL for reservoir simulation that compiles to open-source simulators. Named, validated parameters replace positional keyword decks, and physics checks run before compilation finishes.

4. RIGOR, an open yardstick

   A benchmark for reservoir-engineering agents: hundreds of bespoke tasks in three tiers, from coreflood displacement to mid-conversation compositional pivots, scored from simulator diagnostics.

5. 0.82 against 0.75

   CLARISSA's score, against a coding agent given the full simulator manual. The same model underneath both.

6. SPE-234136-MS

   A Conversational User Interface for Autonomous Reservoir Simulation Deck Generation and Execution. Matejka, Laube, Perschke. Presented at SPE ATCE, Houston, 21–23 October 2026, and at the SPE Summit on AI in Kuala Lumpur.

7. Checked against the classics

   The SPE1, SPE5 and SPE9 comparative cases, generated from their published specifications, ran without error and matched the published results.

8. cap rock, sealing the trap

fences, not rails

the deck must close its material balance

1. Texas A&M University

   B.S. in Nuclear Engineering, 2022. College Station, Texas.

2. Further coursework

   Machine Learning (Stanford); CNNs and sequence models (DeepLearning.AI); Neural Networks: Zero to Hero (Karpathy); mechanistic interpretability (Neel Nanda); intermediate C++.

3. Cloud certifications

   AWS Cloud Practitioner. Google Cloud Associate Cloud Engineer.

4. containment, after the Pantheon
5. control rods
6. fuel assemblies
7. steam generator

College Station, 2022

first principles

1. Image-to-3D benchmarks

   Applied Research Solutions, 2022–2025. An automated benchmark for image-to-3D models (TripoSR, InstantMesh, Trellis, Hunyuan 3D) using Hausdorff and Chamfer distances, plus a vision-language review of 360° turntable video that caught 94% more low-quality meshes than the geometric metrics alone.

2. Satellite segmentation

   Segment Anything fine-tuned on mixed synthetic and real satellite imagery for dish mensuration, reaching 0.89 F1. A distilled MobileSAM in a multi-head network matched it at 5× the speed.

3. Models in service

   Open-weight language models deployed for internal use, agentic RAG over company documents, and a Llama 3.3 fine-tune (DPO) on server logs for outlier detection.

4. 1.5 billion records

   DecisionLinks, 2025. A distributed Milvus cluster on AWS EKS. 256-d Matryoshka embeddings with 8-bit product quantization cut storage 88% for under 1% recall loss.

5. Reading between the lines

   Sparse records enriched with DSPy and GRPO lifted exact-match recall 7%. A config-driven embedding pipeline cut new-dataset onboarding from three days to four hours.

6. Synthetic data at scale

   One million synthetic records from a 24B model across eight H200s in two hours, with under 0.1% duplication.

one and a half billion records

measure the mesh, then watch it turn

1. Tracking guardrails

   An exponential moving average and an attention-inspired template cache let MixFormerV2-Small outperform MixFormerV2-Base.

2. 100% coverage, 6× faster

   Full tracking coverage on the test set across grass, sky and dense forest, where the base model had failed, at six times the inference speed.

3. A custom detector

   A GhostNetV2 detector trained to replace the off-the-shelf models: mAP50–95 of 0.841 and recall at best F1 of 0.873, ahead of the fine-tuned YOLO and RT-DETR baselines.

4. Camera drivers

   C++ drivers for Lucid and Basler GigE cameras, wired into the real-time detection pipeline.

5. objective lens group
6. two-axis gimbal

grass, sky and forest

never lose the target

1. TESSERACT

   The Elder Scrolls System & Environment for Real-time Autonomous Character Technology: a C++ SKSE64 plugin on CommonLibSSE-NG that gives Skyrim's characters language-model agency. They perceive game state, reason, and answer in context. In active development; its moon is a CommonLibSSE-NG plugin template for xmake.

2. Sentient Social Network

   An attempt at a sentient social network modelled on Daniel Dennett's "fame in the brain" theory, built on a multi-agent architecture with Flask, MongoDB and Socket.IO.

3. LANNS

   An implementation of LANNS (VLDB 2022), web-scale approximate nearest-neighbour search that shards and segments data over HNSW, with a pipeline from raw records to a queryable index.

4. Drone Hunter

   A PPO agent that learns to shoot down incoming drones from what a camera can see: NanoDet detections, Kalman-filtered depth and velocity, Hungarian-matched tracks. It scores +91.66, above human play (about 60) and the oracle baseline (+80.9). Runs on Android through Termux.

5. lf-tracking

   Real-time tracking of a 73 × 46-pixel drone in 1080p footage: 717 of 717 frames at 125.8 FPS with MixFormerV2 on ONNX Runtime, using motion compensation, edge-based verification and a recovery search, with no fine-tuning.

seven hundred seventeen of seven hundred seventeen

the innermost began first