Ample News: from grounded research to published media

What workflow did Ample News address?

PR content crosses several disconnected activities: understanding a company, following relevant news, preparing an interview, recording it, editing the result, adapting it for each social platform, and tracking publication. The engineering challenge was to make those stages behave as one inspectable product rather than a collection of AI calls and third-party APIs.

Ample News combined a daily research and interview-preparation workflow with a browser-based voice interview, video and caption processing, clip editing, and platform-aware social distribution. Subscriptions, scheduling, analytics, and observability supported the same operating path.

What was my role and attribution boundary?

I contributed as a full-stack engineer on Ample News through HustleWing. My work covered production AI workflows, multi-provider model boundaries, agent behavior, voice and media integrations, social scheduling, billing, and observability across the product.

This case study describes source-supported contributions and the reviewed system architecture. It does not present Ample News as a separate employer, claim sole authorship of every agent or workflow in the repository, or attribute the product's commercial outcome to engineering alone.

How did the AI workflow remain structured?

The Daily Brief moved through explicit, typed stages for run creation, company and customer context, grounded news search, topic generation, interview preparation, email composition, persistence, and delivery. Each stage had bounded inputs and outputs, contextual logging, and an explicit failure boundary.

• Provider selection kept Gemini, xAI/Grok, and OpenAI behind task-oriented service boundaries
• Tracked run records made long-running work inspectable beyond one request lifecycle
• Structured schemas constrained workflow handoffs and persisted output
• Concurrency limits, retry behavior, daily gates, and reuse checks bounded batch ingestion
• Per-profile isolation prevented one failure from invalidating an entire batch

How were agent actions made safer?

The PR assistant routed user intent to specialized behavior for general support, interest updates, and interview editing. Read operations could happen directly, while state-changing tools required confirmation before writing.

Persistent memory carried relevant preferences between sessions, but tool access remained explicit. The read, confirm, write sequence made mutations understandable to the user and easier to trace when behavior did not match the request.

How did the media and publishing stages work as one system?

Voice interviews, transcript capture, generated hooks, video storage, captions, clip editing, social scheduling, billing, and analytics crossed providers with different delivery and failure models. The product used provider boundaries, webhooks, cron jobs, persisted status, and recovery paths to coordinate them.

• Retell AI voice interviews used a provider abstraction and an XState recording workflow
• Mux and ZapCap supported video playback, caption generation, and webhook-driven processing
• Ayrshare publishing used one first-class social-post record per platform, date, profile, and asset
• Explicit planned, scheduled, posting, posted, and failed states made platform delivery recoverable
• Stripe-backed access rules connected subscriptions to product behavior

How was AI behavior made observable?

Langfuse and OpenTelemetry connected model calls, Mastra workflows, agents, tools, tokens, latency, and cost to the surrounding product context. Structured application logs and Sentry completed the operational view outside the model boundary.

Tracing volume was treated as an engineering constraint. Low-value infrastructure spans could be filtered while preserving the agent, workflow, tool, and model events needed to investigate behavior and cost.

What does the completed project demonstrate?

Ample News demonstrates production AI as workflow engineering: models operated inside typed stages, user confirmations, persistent state, external-service boundaries, recovery paths, and observable operations. The result was a delivered product capability spanning research through distribution, not an isolated chatbot demonstration.

The product was later discontinued after it did not achieve sufficient traction. This case study therefore makes no commercial-success claim and withholds the unverified estimate of daily time saved. Its evidence is the implemented system, the attributable contribution, and the engineering decisions that made a complex AI and media workflow operable.