Constiertel
Consti Ertel
Equity research, analytical systems, and machine learning projects.
I build ML systems, publish equity research, and turn messy datasets into decisions worth making.
Current focus
- Self-supervised vision and generative model research — building toward deployable ML systems.
- Equity research with a long-horizon, cash-flow-first lens focused on durable business models.
- Data science and ML coursework at UC Berkeley spanning statistics, optimization, and inference.
Recent research
Editorial stock workups built for patient readers.
Selected builds
Machine learning and analytics projects with production intent.
JEPA Experiment
A PyTorch research comparison of two self-supervised vision objectives: JEPA latent prediction versus masked-patch reconstruction (MAE-style). Uses a small Vision Transformer on CIFAR-10 with evaluation across linear probing, retrieval, anomaly detection, and embedding visualization — finding JEPA yields more perturbation-invariant representations while MAE performs stronger on linear probe accuracy.
Personalized Listing Photo Editing
A product-oriented ML prototype for personalized real-estate photo editing. Uses BLIP-2 for style discovery, FLUX.1 Kontext as a teacher model to generate content-preserving edits, and InstructPix2Pix with LoRA for student distillation — resulting in compact, photographer-specific adapters (~3 MB each) that reproduce a photographer's signature finish without running a heavyweight model at inference time.
Tesla Manufacturing Cost Simulator
A Bayesian simulation (Project Aurora) comparing US, Mexico, and China manufacturing costs across raw materials, labor, logistics, FX, tariffs, and discrete risk events. Replaces hand-crafted Normal priors with data-driven Student-t and Beta posteriors fit to FRED economic series, producing wider, more realistic uncertainty intervals across all factory scenarios.