AI Research · Computer Vision · 3D Scene Understanding
A Pennsylvania-based AI research company advancing self-supervised visual learning for remote sensing, feed-forward 3D scene reconstruction, and secure intelligent systems — where data finds its flow.
Six Rivers LLC is a Pennsylvania-based AI research company exploring the frontiers of computer vision, deep learning, and intelligent systems.
Our name reflects a philosophy: insight emerges at the convergence of many streams of inquiry. Like rivers flowing together, we bring together computational theory, model architecture, data engineering, and real-world application to produce meaningful research.
Our research spans three core directions: self-supervised visual learning for remote sensing, feed-forward 3D reconstruction, and emerging work in AI security.
Four converging programs of inquiry — from satellite imagery to real-time 3D scenes to secure AI infrastructure to AI-assisted mathematics.
Applying the DINO family of self-supervised vision transformers (DINOv1 → DINOv2 → DINOv3) to satellite and aerial imagery. We extract robust visual features without labeled data, perform unsupervised object discovery and semantic segmentation via attention maps and PCA-based foreground segmentation, and benchmark against traditional remote sensing methods.
Specific investigations include: domain transfer from natural images to overhead scenes using k-NN classification, attention-map object localization in satellite imagery, DINOv3 Gram Anchoring for preserving dense feature quality, dense feature matching for multi-temporal change detection, and lightweight modality adapters for multi-spectral bands (Sentinel-2 NIR/SWIR, Landsat-8).
Advancing real-time 3D reconstruction via feed-forward Gaussian Splatting, building on the Szymanowicz progression: Splatter Image (single-view object reconstruction at 588 FPS), Flash3D (monocular scene reconstruction with depth foundation models), and Bolt3D (generative diffusion-based multi-view synthesis).
We investigate cross-domain transfer to aerial and remote sensing imagery via SkySplat (first generalizable 3DGS for satellite NVS), pose-free reconstruction with NoPoSplat, voxel-aligned Gaussian prediction (VolSplat) pushing multi-view NVS beyond 31 dB PSNR, and scene token fusion (GlobalSplat) for unbounded scene reconstruction. Our work leverages VGGT — the CVPR 2025 Best Paper — for joint camera-depth-pointmap estimation.
Exploring the intersection of AI and cybersecurity: agent auditing and safety evaluation, model robustness against adversarial inputs, secure deployment infrastructure for autonomous systems, and audit frameworks for LLM-based agent behavior in production environments.
A parallel thread investigates AI-assisted computation in pure and applied mathematics: Grobner basis computation via learned algebraic reasoning, VC dimension estimation for hypothesis classes, and Hamming-distance-constrained search over structured combinatorial spaces. These directions aim to accelerate traditionally symbolic or combinatorial workloads through learned approximations and differentiable relaxations.
We also explore AI-powered cryptanalysis, including neural approximations for lattice-based hardness assumptions such as the Short Integer Solution (SIS) problem and Learning With Errors (LWE), with the goal of understanding when learned attacks can outperform classical reductions and where structural invariants resist current learned approaches.
Complementary applications include automated literature synthesis and AI-assisted assessment — tools that extend the reach of mathematical and technical research. Research questions and partnerships across these directions are being established for late 2026 and beyond.
Our approach is hypothesis-driven, reproducible, and grounded in state-of-the-art literature.
Multi-spectral satellite imagery preprocessing (Sentinel-2, Landsat-8) with radiometric correction, cloud masking, and ground-sample-distance encoding. Self-supervised pretraining with multi-crop augmentation (global + local crops), followed by unsupervised object discovery via attention maps and PCA-based foreground segmentation.
Feed-forward regression and generative diffusion for novel view synthesis, from single objects to unbounded scenes. Depth foundation models (UniDepth, Depth Anything V2) enable cross-domain generalisation. Voxel-grid aggregation (VolSplat) and scene token fusion (GlobalSplat) reduce the computational graph to milliseconds per scene.
Agent auditing frameworks for evaluating safety and robustness of autonomous LLM-based systems in production environments. Exploration of neural cryptanalysis for lattice-based primitives (SIS, LWE). AI-assisted algebraic computation (Grobner basis methods), statistical learning theory (VC dimension estimation), and combinatorial optimization over Hamming spaces.
Bringing research to practice — products and services developed from our core research programs.
In Development
A scalable pipeline for self-supervised analysis of satellite and aerial imagery. Leverages DINOv2/v3 vision transformers to perform object discovery, semantic segmentation, and change detection without pixel-level labels. Designed for precision agriculture, infrastructure monitoring, and environmental assessment — where labeled training data is scarce but raw imagery is abundant.
In Development
A feed-forward Gaussian Splatting engine for producing 3D scene reconstructions from monocular or sparse multi-view inputs. Built on Splatter Image / Flash3D / Bolt3D and extended with VolSplat voxel aggregation and GlobalSplat scene token fusion. Targets real-time performance (sub-second per scene) for surveying, construction monitoring, and aerial remote sensing applications.
Research Phase
Frameworks and benchmarks for auditing the safety, robustness, and reliability of autonomous LLM-based agents in production. Covers adversarial input testing, jailbreak detection, behavioral consistency evaluation, and prompt-injection defense validation. Targets organizations deploying AI agents in customer-facing or sensitive-system contexts.
Research Phase
Accelerating traditionally symbolic and combinatorial workloads through learned approximations: Grobner basis computation via learned algebraic reasoning, VC dimension estimation for hypothesis classes, and neural cryptanalysis of lattice-based primitives (SIS, LWE). Targets researchers and practitioners bridging machine learning with pure and applied mathematics.
For inquiries about research collaboration, compute partnerships, or general questions — we'd be glad to hear from you.