Text & long-form
On-brand articles, RFP responses, knowledge-base content. Style transfer trained on your past best work.
LoRA · style-fitAI / ML · 02 of 05
Generation, on your brand. At your scale.
Text, code, image and multimodal pipelines that produce output worth shipping. We fine-tune to your voice, ground to your sources, and bolt on an evaluation harness so "it sounded fine" never has to be the QA answer again.
140kGenerations / day · top client
78%Pass eval on first try
−61%Cost vs. naive GPT-4
promptgenerate(brand="orbital", modality="mixed")
// onboarding email
export const mailer = async () => {
const tpl = await render("welcome", user);
return send(tpl);
}
export const mailer = async () => {
const tpl = await render("welcome", user);
return send(tpl);
}
BLEU 0.41brand-fit 9.2/10$0.003 / gen
What you get
Six generative capabilities, one pipeline.
We pick the smallest model that does the job, ground it in your data, and ship it with evals you can read in a meeting.
Code generation
In-editor copilots, scaffolding agents, code review bots — tuned to your repo conventions and your test suite.
repo-aware · typedImage & product
SDXL / FLUX pipelines for catalog, lifestyle and creative variants — consistency across generations.
SDXL · FLUX · IP-adapterVideo & motion
Short-form social, product demos, training reels with Runway and Sora-class APIs.
Runway · Sora · ComfyUIRetrieval-augmented
Hybrid (BM25 + vector) retrieval over your docs, tickets and tables. Citations on every answer.
hybrid · rerank · citeEvaluation harness
Per-task golden sets, model-graded rubrics, regression alerts in CI.
golden set · rubric · CIHow it works
Ground first, generate second, evaluate always.
A three-lane pipeline that's survived the boring middle of dozens of generative projects.
01 · Ground
DocsTicketsBrand assetsEmbeddings
02 · Generate
Base LLMLoRA / SFTToolsConstraints
03 · Evaluate
Golden setRubricHuman reviewRegression CI
The first 70% is grounding. The last 30% is taste.
Most "generative AI" failures aren't model failures — they're grounding failures. We start with the documents, examples and constraints that anchor output to your reality.
- 01
Curate the source-of-truth
Ingest pipelines, chunking strategy, freshness rules — the boring part that determines quality.
- 02
Pick the smallest model that wins
Frontier models for hard tasks, distilled SLMs for the easy 80%.
- 03
Fine-tune only what's worth it
LoRA / SFT for style and structured output.
- 04
Eval is a CI gate, not a milestone
No model ships if regression fails.
Tech stack
Picks we'd defend at 3am.
Generative tooling moves weekly. We track it so you don't have to — and we only ship what survives a quarter of production traffic.
Foundation models
GPT-4o / o1Claude 3.5Gemini 1.5Llama 3.1Mistral Large
Image & video
SDXLFLUXDALL-E 3Stable VideoRunway Gen-3ComfyUI
Fine-tuning
AxolotlUnslothHF PEFTDPO / ORPOLoRA / QLoRA
Retrieval
pgvectorWeaviateVespaCohere RerankBM25
Orchestration
LlamaIndexLangChainDSPyInstructor
Eval
BraintrustLangSmithPromptfooRagas
Serving
vLLMTGIModalReplicateBedrock
Workflow
TemporalAirflowPrefectArgo
From vision to victory
From prompt to production pipeline, in five steps.
No 9-month transformations. A real generative feature, in production, with eval gates — in six weeks.
01
Week 1Sharpen the brief
Pick a single generation task. Define what good output looks like using a rubric.
02
Week 2Build the golden set
30–80 input/output examples. This becomes your permanent quality anchor.
03
Week 3–4Generate & ground
RAG pipeline + prompt structure + optional LoRA. Iterate against dataset, not opinions.
04
Week 5Ship behind eval gate
CI blocks regressions. Cost and latency budgets enforced automatically.
05
OngoingImprove continuously
Production failures feed dataset. Improvements compound over time.
Where it pays back
Three pipelines we've scaled.
Patterns we've run hot — with the cost, quality and throughput curves to prove it.
A content factory that sounds human.
120k SKUs × 7 locales of product copy. LoRA + retrieval + reviewer queue.
−84%Cost / SKU
9.2 / 10Brand-fit score
Claude 3.5Llama LoRApgvector
A copilot that knows your repo.
Repo-aware assistant grounded in codebase, conventions and test suite.
+38%PRs merged in 1 review
−27%Bug regressions
GPT-4oInstructorTree-sitter
Lifestyle imagery without a studio day.
SDXL + IP-adapter pipeline for furniture catalog scaling.
32 hr → 11 minPer SKU
1.4MImages / month
SDXLIP-adapterComfyUIModal
Why ETY
Pipelines, not parlor tricks.
2.1MGenerations / day across all client pipelines we operate.
78%Average eval-pass rate across our generative deployments.
11Models we've fine-tuned & deployed in production over 18 months.
−61%Median cost reduction vs. baseline frontier-only setup.
One pipeline. Receipts on the rubric.
Send us a single generation task that's slow, costly or off-brand today. We'll come back with the golden set, the pipeline and the eval gate — or tell you the smaller fix that solves it cheaper.