Private Mac notes for data scientists

Data scientists, machine learning engineers, and quantitative researchers operate at the intersection of mathematical theory, statistical experimentation, and production software engineering. In an active machine learning lifecycle, a data scientist tests multiple model architectures, tunes hyperparameters, evaluates cross-validation loss curves, and conducts collaborative review calls with data engineering and product stakeholders.

Documenting machine learning experimentation requires precision and reproducibility: recording the exact learning rate schedule, tracking data cleaning assumptions, noting benchmark inference latencies, and documenting why a specific loss function was chosen over another.

Yet mainstream note-taking platforms are fundamentally ill-suited for scientific computing. Web-based document tools fail to render monospace code properly, mangle LaTeX equations, and lock experimental findings inside proprietary cloud silos. Furthermore, using commercial AI meeting tools to record algorithm reviews introduces noisy bots that distract researchers and upload unreleased model architectures to third-party cloud servers.

Oyma is the dedicated research workstation engineered for data scientists on Apple Silicon Macs. Operating locally on your hardware, Oyma captures model review calls silently via ScreenCaptureKit, transcribes technical speech 100% on-device with local Whisper models, and organizes your machine learning experiments in an interconnected Markdown knowledge base.

The experimental reproducibility gap in machine learning

Data science teams regularly struggle with knowledge fragmentation across the model development lifecycle:

1. Undocumented hyperparameter decisions: Informal discussions about why a learning rate was adjusted or why certain outlier features were dropped are verbalized on calls but never documented in the codebase.
2. Bot friction during research reviews: Inviting an automated bot into an exploratory modeling session distracts researchers and triggers concerns about intellectual property exposure.
3. Proprietary algorithm leaks: Model architectures, prompt engineering templates, and training datasets represent core competitive advantages. Sending audio describing these assets to cloud transcription vendors creates severe legal and IP risks.
4. Disconnection from code and data: Cloud note platforms cannot be parsed by command-line utilities, version-controlled with Git, or linked directly to repository directories.

The data scientist’s experiment knowledge graph

Oyma organizes your modeling workflows into an interconnected graph of experiments, benchmark evaluations, and architectural decisions:

┌────────────────────────────────────────────────────────────────────────┐
│                   DATA SCIENCE EXPERIMENT TOPOLOGY                     │
├────────────────────────────────────────────────────────────────────────┤
│                                                                        │
│                    [[Model: Embedding Distillation]]                   │
│                    models/embedding-distillation-v2.md                 │
│                                ▲                                       │
│          ┌─────────────────────┼─────────────────────┐                 │
│          │                     │                     │                 │
│          ▼                     ▼                     ▼                 │
│  [Experiment Run #14]   [Experiment Run #18]   [Model Review Session]  │
│  Loss: Cosine similarity Loss: Triplet ranking Evaluation with ML Lead│
│  "Batch size 512"       "Temperature 0.07"     "Approve production PR" │
│          │                     │                     │                 │
│          └─────────────────────┼─────────────────────┘                 │
│                                ▼                                       │
│                  [[Production Benchmark: Latency]]                     │
│                  benchmarks/inference-p99-latency.md                   │
│                                                                        │
└────────────────────────────────────────────────────────────────────────┘

Using bidirectional [[wikilinks]], each experiment log connects directly to the core model architecture dossier. When presenting findings to leadership or onboarding a new research engineer, you can trace the entire lineage of modeling decisions in seconds.

Research synthesis loop: From algorithm brainstorming to verified experiment

1. Botless loopback during technical reviews

When you pull up Jupyter notebooks or Weights & Biases dashboards during a team review on Zoom or Meet, Oyma captures loopback audio via macOS CoreAudio. No robotic guest tile joins the call, avoiding distractions during delicate statistical debates.

2. Integrated mathematical scratchpad

As team members discuss loss surface geometry or batch convergence rates, write live LaTeX equations, embed Python code blocks, and sketch hyperparameter matrices directly in Oyma‘s distraction-free editor.

3. Local neural transcription of ML terminology

The moment the debrief ends, Apple Silicon’s hardware matrix units transcribe the audio using Whisper. The model recognizes machine learning terms like InfoNCE, AdamW, cosine annealing, and KL-divergence accurately without transmitting data to remote cloud clusters.

4. Git-trackable Markdown experiment logs

The resulting document saves directly into your research repository as plain .md. You can version-control notes alongside PyTorch code, search historical runs with terminal tools, or navigate model relationships in Obsidian.

Authentic machine learning experiment note sample

Here is a realistic data science experiment note produced in Oyma:

---
title: "Experiment Log: Contrastive Learning Fine-Tuning (Run 18)"
date: 2026-10-08
model: "bge-micro-v2"
framework: PyTorch
dataset: "enterprise-docs-corpus-v4"
tags: [data-science, embeddings, contrastive-learning, pytorch, evaluation]
---

## Objective & Theoretical Hypothesis
Fine-tune a 128-dimensional embedding model using InfoNCE loss with hard negative mining. Hypothesize that dynamic temperature scaling will improve top-5 retrieval recall on technical documentation queries without increasing embedding vector dimensionality.

## Hyperparameter Configurations & Training Snippet
```python
import torch
import torch.nn.functional as F

def info_nce_loss(query_embeds, doc_embeds, temperature=0.07):
    # Normalized cosine similarity matrix
    q_norm = F.normalize(query_embeds, p=2, dim=1)
    d_norm = F.normalize(doc_embeds, p=2, dim=1)
    sim_matrix = torch.matmul(q_norm, d_norm.T) / temperature
    labels = torch.arange(query_embeds.size(0), device=query_embeds.device)
    return F.cross_entropy(sim_matrix, labels)

Validation Benchmark Evaluation

Run ID	Loss Function	Temperature	Top-1 Recall	Top-5 Recall	P99 Latency
Run-12	Triplet Margin	N/A	68.4%	84.1%	1.8ms
Run-16	InfoNCE	0.10	72.1%	88.5%	1.9ms
Run-18	InfoNCE (Hard Neg)	0.07	76.8%	92.4%	1.9ms

Model Review Decisions & Consensus

• Approved Run-18 weights for staging deployment following review with ML platform lead.
• Agreed to maintain a 128-dimensional projection layer to keep memory footprint under 256MB on edge inference nodes.

Action Items

• Data Scientist: Export quantized ONNX model artifacts to the model registry
• ML Engineer: Benchmark batch inference throughput on Apple Silicon M-series hardware
• Team: Prepare A/B evaluation plan against current production baseline

## Absolute protection for model weights and training datasets

Data scientists and ML researchers handle highly proprietary algorithmic designs, training data distributions, and competitive model benchmarks. Transmitting verbal deliberations about model architectures to external cloud recording vendors creates massive trade-secret exposure.

Oyma secures your research pipeline:
- **Complete on-device execution:** Acoustic models and Markdown generation operate strictly on Apple Silicon without third-party network APIs.
- **Inspectable packet containment:** Use the native [Egress Log](/features/egress-log) to verify that zero telemetry or audio fragments leave your Mac.
- **Interoperable file formats:** Every experiment log lives as open Markdown on your FileVault-encrypted volume, accessible to Python scripts, command-line pipelines, and local grep.

Advance your machine learning breakthroughs with reproducible research logs, botless review capture, and private Markdown notes on your Mac.