Research
We publish what we learn building production email infrastructure on AWS Bedrock. The goal isn't novelty — it's writing down the answers we wish someone else had written down for us. Three areas of focus.
1. Intent inference from open signals
How much of "is this prospect going to convert?" can be predicted from public signals alone — site content, product catalog, hiring posts, blog cadence — without enrichment from third-party data brokers? We've found you can hit AUC ≈ 0.79 on a Shopify-merchant population using only public signals scraped at indexing time. That's commercially meaningful.
Read the deep-dive on the 40 features that drive most of the score.
2. LLM personalization at production scale
The naive prompt template that works on 10 generations breaks at 10,000. We track failure modes: hallucinated product names, brand-voice drift, repetition across a list, the LLM catastrophically misreading a niche. Defending against each requires its own evaluator, run on every output before it goes to SES.
Current eval stack runs in-process: a fact-check pass against the structured prospect record, a tone-classifier, a similarity check against the last 200 outputs from the same campaign. Together they catch ~94% of issues before send. The remaining 6% gets human review on the Growth plan.
3. Deliverability as an engineering problem
Deliverability isn't a marketing problem; it's an infrastructure problem with a marketing UI. We treat it the same way we'd treat database replication lag: instrumented, alerting, runbook'd. Postmaster Tools complaint rate, SNDS reputation, and our own ingest of bounce-NDR codes feed a weighted health score per sending domain. Below 0.85 we automatically throttle.
The 2025 warmup pattern we use on every new domain.
What we don't research
We don't publish thinly-sourced "AI trends" content. We don't run pre-print papers. The goal is to be a useful resource for someone building a similar system, not to compete on academic novelty. If we discover something genuinely new, the right place for it is a paper, and we'll do the separate work to publish it properly.
Collaboration
If you're working on adjacent problems and want to compare notes, write to pierre@parisai.click. We're particularly interested in anyone working on cross-tenant abuse signal sharing for cold email — that one's hard and we don't have a clean answer yet.