AI & Automation

The hard part of shipping AI is not calling the model. It is what the system does when the model is wrong, when it is expensive, and when what it produces will be read by someone who will act on it.

How we actually work

The model may not invent facts. Clepora Jobs reorders, reframes and rewords what a candidate actually did. It never adds an employer, a date or a skill. Where a real gap exists it surfaces in missing_keywords with an honest recommendation instead of being papered over. This is a hard guardrail in the system prompt and a tested property — not a hope. It also happens to be the commercially correct choice: fabrication on a résumé gets people fired, and a product that does it to them is a liability, not a feature.

Cheap deterministic filter, expensive model only on finalists. Clepora Jobs scores every posting with a fast, explainable blend of skill coverage (75%) and title match (25%), then spends model tokens only on what clears a ≥75 gate. The scalable pre-filter and the nuanced judgement are different tools; using the second for the first is how AI products become unaffordable at the exact moment they start working.

Structured output, validated on the way in. Tailored résumés and match analyses come back as schema-constrained JSON parsed straight into Pydantic models. Where the structured-output API rejects a JSON-Schema keyword, we strip it before sending and Pydantic re-applies it on validation, so no constraint is silently lost.

The provider is a slot, not a dependency. Every AI system we build ships with an adapter — Anthropic direct, AWS Bedrock, and a keyless mock that makes the whole product work end to end with no credentials and no spend. Newsroom’s full feature set runs on mock; only the prose quality changes. That is what lets a test suite of 715 tests run without a key, and what stops a provider outage or a price change from being an architecture problem.

Spend is bounded and never automatic. Newsroom’s scheduler will build a bulletin but will not render audio — that costs money, so it waits for a human to press the button. The paid video-avatar path carries an explicit spend cap. Every scheduled run is recorded and surfaced at /api/health, so a job that has been quietly failing is visible rather than invisible.

Synthetic media, handled responsibly. Newsroom’s presenter is drawn, not photoreal, and carries a permanent non-dismissible “AI PRESENTER” mark. A photorealistic synthetic anchor reading real news is the precise format used in disinformation campaigns; we declined to build it. Story photographs are the publisher’s own, credited and linked — never generated.

What that produces

Engage us for

• AI feature build — a model-backed capability inside your existing product, with the guardrails and the eval.
• Cost & reliability review — an AI system that works but costs too much or fails opaquely.
• Responsible-AI review — where your synthetic content, automation or decisioning creates exposure you have not priced.