AI-Driven Email Personalization at Scale

Cold email at volume dies on generic copy. Open rates for bulk outreach average 2 to 4 percent in B2B SaaS. The message looks templated because it is. Recipients recognize the pattern in the first sentence and delete without reading.

The Avo Engine needed to send cold outreach to hundreds of marketing agency leads without paying for Apollo, Clearbit, or any managed enrichment provider. The build-don't-rent mandate ruled out every SaaS email tool that charged per record. The goal: generate genuinely personalized pitches from public information alone, at 500-lead batch scale, for a cost approaching zero.

The constraint that defined the architecture: each pitch had to read as if it were written by someone who spent ten minutes on the prospect's website. Not mail-merge. Not persona slots. Actual specific language pulled from what the company actually says about itself.

The pipeline had three stages: fetch, extract, generate. For each lead record, a Rust scraper issued an HTTP GET to the prospect's primary domain with a 10-second timeout and a user-agent that matched a standard browser. No JavaScript rendering. Pure HTML. The raw response was stripped to text using a lightweight HTML-to-text parser that preserved semantic structure while discarding navigation chrome and footer noise.

The stripped text was scanned for key phrases: service offerings, stated client types, outcome language, technology mentions, and social proof signals. This extraction ran as a Rust function with zero API calls. The output was a structured JSON object with offerings, client_types, outcome_language, and social_proof fields.

The structured JSON was passed to Claude Haiku with a prompt template that constrained output length to 80 to 120 words, tone to direct engineer-to-engineer, and structure to a 1-sentence hook referencing something specific from their site, a 2-sentence problem statement, a 1-sentence offer, and a CTA. No dashes, no filler phrases, no generic opener.

Confidence scoring was binary. If the fetch succeeded and extraction returned at least two populated fields, the record got confidence 0.75. If the fetch failed or extraction returned empty, the record fell to confidence 0.50 and received a category-level template.

The batch processor was a single Rust binary that read leads from a ClickHouse table, fetched and scored each one concurrently using Tokio with a semaphore limiting concurrency to 50 simultaneous HTTP connections, and wrote results to alien.personalized_emails (ReplacingMergeTree keyed on lead ID).

The Claude API call used streaming disabled, max tokens 200, temperature 0.4. Temperature below 0.5 kept output consistent across similar inputs without making the copy robotic. The prompt included five few-shot examples of accepted outputs and two examples of rejected outputs with inline labels explaining why they failed.

One postmortem finding: the first version of the extraction step used regex to pull service descriptions. On sites with heavy JavaScript-rendered content, the HTML contained almost nothing. The fix added a secondary extraction pass using Claude Haiku on raw HTML snippets below 2,000 tokens for fallback parsing, at a cost of roughly $0.0003 per call. This fallback path recovered 23% of records that would otherwise have been templated.

Output was stored as drafts only. No send was triggered automatically. The store-as-draft policy was non-negotiable and enforced at the schema level: the personalized_emails table had no sent_at column.

500 leads processed. 353 records reached high-confidence status. 147 received category-level templates. Total Claude API cost for the batch: approximately $1.40. Processing time for 500 leads with 50-concurrent fetch: 4 minutes 12 seconds.

In early A/B testing on the send queue, personalized drafts showed a 34% open rate versus 11% for the category-level templates. Click-through on the personalized variant ran at 6.2% versus 1.8% on templates. The filler-phrase filter prevented an estimated 67 low-quality personalized pitches from reaching the send queue.