DoublewordDoubleword

Workbooks

Production-ready examples demonstrating what becomes possible when LLM inference is cheap enough to use at scale. Each example includes working code, real-world data, and measured costs.

To get started, install the dw CLI and log in:

curl -fsSL https://raw.githubusercontent.com/doublewordai/dw/main/install.sh | sh
dw login

Or sign up at app.doubleword.ai if you don't have an account yet.

Examples

ExampleWhat It DoesCostKey Insight
async-agentsDeep research with recursive agent trees$0.34 for 47 agents1-hour SLA enables multi-round agentic workflows
synthetic-data-generationGenerate training data with quality filtering$3.21 for 10K samples3-stage pipeline in 3 hours, not 3 days
data-processing-pipelinesClean and enrich messy records$0.80 for 50K recordsLLM-powered normalization at pipeline scale
embeddingsSemantic search over document corpus$0.03 for 1.6M tokens70% cheaper than OpenAI for same quality
model-evalsBenchmark models on GSM8K$0.21 for 1,319 questionsComprehensive evaluation becomes routine
bug-detection-ensembleClassify security vulnerabilities$0.40 for 4,642 samplesRun twice for calibration, still under $1
dataset-compilationBuild company datasets via search + LLM$1.05 for 188 companies100% recall vs Gartner Magic Quadrant
structured-extractionExtract fields from scanned receipts$0.12 for 626 receiptsQwen3-30B beats GPT-5.2 at 1/12th cost
image-summarizationCaption images for social media$0.10 for 1,000 imagesVision batch makes captioning automatic

Why These Examples Matter

Each example demonstrates a "more is different" capability: something that becomes qualitatively new when inference costs drop by 10-50x.

Multi-stage pipelines benefit from the 1-hour SLA. Synthetic data generation and data processing both run three sequential batches. With a 24-hour SLA, that's three days minimum. With Doubleword's 1-hour SLA, the same pipelines complete in 3 hours. This changes iteration speed: you can refine prompts and re-run the full pipeline multiple times in a single day.

Agentic workflows compound the SLA advantage. The async-agents example spawns recursive agent trees that require 20+ batch rounds. At 24 hours per round, that's potentially a month of wall-clock time. At 1 hour per round, it completes in a day.

Single-batch workloads are about cost. Embeddings, model evals, and image summarization process everything in one batch. The 24-hour SLA is fine; what matters is the 70-95% cost reduction versus real-time APIs. At these prices, you can embed your entire corpus, evaluate every model on every benchmark, and caption your whole image library.

Cost Comparison

Across all examples, Doubleword's batch pricing delivers 10-50x cost savings versus real-time APIs from OpenAI and Anthropic:

TaskDoublewordOpenAI GPT-4oAnthropic Sonnet 4.5Savings
Deep research (2M tokens)$0.34$5.81$7.2517-21x
Synthetic data (20M tokens)$3.21$108.83$154.6234-48x
Data cleaning (6M tokens)$0.80$27.40$38.1534-48x
Model evaluation (400K tokens)$0.21$1.06-5x
Document embeddings (1.6M tokens)$0.03$0.10-3x

Pricing sources: Doubleword, OpenAI, Anthropic.

Running the Examples

Each example is a dw project — a self-contained workflow defined in a dw.toml manifest. Clone any example and run it with the dw CLI:

dw examples clone model-evals
cd model-evals
dw project setup
dw project info

Run everything end-to-end:

dw project run-all

Or run steps individually for more control — see dw project info for the full workflow, and each example's README for step-by-step instructions.

How dw Projects Work

Each example has a dw.toml that defines:

  • Setup — dependency installation (e.g., uv sync)
  • Custom project steps — Python commands for data preparation and analysis
  • Workflow — the full end-to-end sequence mixing dw CLI commands with project steps

The dw CLI handles batch file management, submission, progress watching, and result retrieval. Your project steps handle domain-specific logic: loading datasets, building prompts, and analyzing results.

dw project run prepare          # Your code: generate batch JSONL
dw files prepare batch.jsonl --model Qwen/Qwen3-VL-30B-A3B-Instruct-FP8
dw batches run batch.jsonl --watch --output-id .batch-id  # CLI: upload, batch, watch
dw batches results --from-file .batch-id -o results/results.jsonl  # CLI: download results
dw project run analyze -- -r results/results.jsonl  # Your code: score/analyze
dw batches analytics --from-file .batch-id  # CLI: cost and token breakdown

Models

Examples use models available on the Doubleword platform:

ModelBest ForBatch Pricing (per 1M tokens)
Qwen3-VL-30B-A3B-Instruct-FP8Best value for most tasksInput $0.05, Output $0.20
Qwen3-VL-235B-A22B-Instruct-FP8Maximum accuracyInput $0.10, Output $0.40
Qwen3-Embedding-8BEmbeddingsInput $0.01

Set the model when preparing batch files:

dw files prepare batch.jsonl --model Qwen/Qwen3-VL-30B-A3B-Instruct-FP8

See the full model list and current pricing at doubleword.ai/pricing or run dw models list.

Project Structure

Each example follows this layout:

example-name/
├── dw.toml            # Project manifest (steps, workflow)
├── README.md          # Results, methodology, replication instructions
├── pyproject.toml     # Python dependencies and CLI entry point
├── src/
│   ├── cli.py         # Click CLI (prepare, analyze, run, etc.)
│   └── ...            # Task-specific modules
└── results/           # Output artifacts (gitignored)

Requirements

  • dw CLI — install with curl -fsSL https://raw.githubusercontent.com/doublewordai/dw/main/install.sh | sh
  • Python 3.11+
  • uv for dependency management (installed automatically by dw project setup if not present)
  • Doubleword account — sign up at app.doubleword.ai
  • Some examples require additional API keys (Serper for web search)
View source on GitHub