Atlantic Companion

Atlantic Companion is an AI assistant built for The Atlantic. It finds articles from the magazine's archive — going back over a hundred years of continuous publication — in response to whatever a reader is curious about right now.

You ask it about anything: a current event, a half-remembered phrase, a person, a question your kid asked you. It returns Atlantic pieces from across the archive that bear on it — sometimes from this month, sometimes from 1932. The retrieval is RAG-backed; the ranking is editorial-curated; the conversational surface is built around the assumption that the reader doesn't already know what they want to read.

It lives at labs.theatlantic.com, behind Atlantic SSO — a logged-in reader product, not an open marketing surface.

Internally, the app shipped from experiments/zeitgeist and was codenamed Abby through development. Atlantic Companion is the public-facing name.

The engagement context

The Atlantic Labs is the wider engagement Atlantic Companion was built inside: a long-horizon partnership between The Atlantic and Metalab to give the publisher a responsible, durable way to explore AI surfaces on top of its own work. Metalab leads it as a full build — product strategy, design, and engineering — and Atlantic Companion is the first reader-facing product to ship under that umbrella.

The brief is wider than any one deliverable: how does a 168-year-old publisher use AI-native surfaces to make its archive more useful — without compromising voice, editorial standards, or reader trust? Atlantic Companion is the most-developed answer so far. Other experiments in the Labs explore adjacent surfaces and form factors.

The retrieval problem

The Atlantic has been publishing continuously since 1857. The archive is the moat — and historically, it has been almost entirely invisible to readers in the moment. Atlantic Companion's job is to make that archive feel like a single, browsable corpus regardless of when a piece was written.

That introduces problems that don't show up in shorter-horizon corpora:

• Vocabulary drift. The language a 1948 piece uses about a topic is not the language a 2024 piece uses for the same topic. Naive embeddings underweight historically important matches because the surface vocabulary has moved on.
• Density. The archive is uneven — eras with heavy coverage of a topic can dominate retrieval to the point that nothing else surfaces. "Relevance" has to be balanced against "what would a curious reader actually want to see across a century".
• Voice. The reader is talking to The Atlantic, not to a generic assistant. Atlantic Companion has to feel like an editor at the magazine, not like a chatbot wrapped in a logo.

The retrieval layer handles drift and density. The editorial-curated ranking and the conversational surface handle voice.

The pattern underneath

What makes the Labs engagement scale beyond a single product — and what I owned as engineering lead — is the operating pattern underneath every experiment:

• Experiment-as-package. A Turborepo monorepo, with each experiment scoped as its own app inside experiments/. Atlantic Companion ships from experiments/zeitgeist. New experiments inherit the toolchain (TypeScript, build, test, deploy) without setup tax — spinning up the next hypothesis costs hours, not weeks.
• One CI/CD pipeline, many surfaces. A single deployment harness handles reader-facing apps, internal admin tools, and backend services. The same pipeline that ships Atlantic Companion ships the editorial dashboard behind it.
• Shared retrieval, swappable ranking. Every experiment talks to the same recommendation backend. Ranking models are swapped behind a model parameter so editorial can A/B which lens works best — Balanced, Timely, Long-form, or anything we add next — without redeploying the app.
• Editorial as a first-class user. Editorial controls are not a backend tuning knob. They're a product surface, named in plain English, used by The Atlantic's own team to steer what the assistant surfaces.
• Privacy by default. Atlantic Companion acts only on what the reader gives it in the conversation. No background classification, no silent profile-building.

Architecture sketch

   ┌─── ATLANTIC SSO ─────────────────────────────────────┐
   │  labs.theatlantic.com · authenticated readers only   │
   └────────────────────────┬─────────────────────────────┘
                            │
   ┌────────────────────────▼─────────────────────────────┐
   │  ATLANTIC COMPANION (experiments/zeitgeist)          │
   │  conversational surface · context-aware retrieval    │
   └────────────────────────┬─────────────────────────────┘
                            │
   ┌────────────────────────▼─────────────────────────────┐
   │  Shared retrieval / ranking service                  │
   │  Atlantic archive index (1857 →) · Balanced /        │
   │  Timely / Long-form / …                              │
   └────────────────────────┬─────────────────────────────┘
                            │
   ┌────────────────────────▼─────────────────────────────┐
   │  EDITORIAL                                           │
   │  ranking-model selector · curation surface · audit   │
   └──────────────────────────────────────────────────────┘

A Terraform-managed AWS footprint underneath. A Turborepo monorepo on top. Atlantic Companion sits on the shared retrieval/ranking service alongside every other experiment in the Labs.

What I owned

• Technical lead across the engagement — not any single experiment.
• The experiment harness: monorepo layout, shared infrastructure, the CI/CD pipeline that lets a new experiment go from git init to deployed in days.
• The contract between editorial, the retrieval/ranking service, and Atlantic Companion's conversational surface.
• AWS infrastructure via Terraform; the deployment story The Atlantic could operate without us.
• The editorial-surface design — translating "what makes a good match for this reader" between domain experts and the ranking team into a named, controllable product surface.
• Atlantic-SSO integration; making the assistant a logged-in reader product on day one.

What carried forward

The most useful pattern from Labs — the one that will outlive any specific experiment — is treating the editorial control surface as a first-class product, not a tuning knob. The moment the model choice (Balanced / Timely / Long-form) was named, surfaced, and controllable, the conversation about quality moved from "is the AI right?" to "which lens do we want for this reader?". A much better conversation to have. A version of this thinking shows up in The Quiet Trade: judgment is the priced layer; the deliverable should make that judgment visible, not hide it.