~ /anthony
[ available ]
// about
~/anthony/README.md
~  $  cat README.md

# anthony carnevale

applied ml engineer · full-stack · turning research-y ideas into production systems

## who_i_am

I build the ML side of products people actually use. The work I enjoy most sits in a narrow slice: LLMs that drive real workflows (not chatbots), embedding-powered search and ranking, and NLP pipelines that take messy human input and return structured, validated data. I'm equally comfortable shipping the FastAPI service behind it, the React widget on top, and the evaluation harness that keeps it honest.

## what_ive_shipped

neptune booking — multi-tenant scheduling SaaS

designer, engineer, founder · 2026

  • - designed an LLM-powered onboarding agent (gpt-4o-mini, json mode) that turns a paragraph of plain text into a working booking config — hours, services, resources, pricing, branding
  • - built semantic service search on top of openai text-embedding-3-small + mongodb atlas vector search, scoped per tenant
  • - shipped an NLP booking parser (microsoft recognizers + custom matching) that handles dates, times, ranges, durations, party size, resource preferences and AM/PM disambiguation
  • - wrote the multi-turn conversational admin agent with tab-aware prompts and a deterministic post-processor that validates every config before merging
  • - stood up the eval harness: 8 business archetypes, structural validators, cross-reference checks, retry-on-truncate logic

data engineering & full-stack — prior roles

  • - architected data pipelines with snowflake, airflow, and dbt
  • - built serverless apis and apps with aws lambda
  • - ran complex database migrations and schema redesigns
  • - shipped full-stack apps with react, node.js, and python

## how_i_think_about_ml

  • - the model is the easy part — the prompt, the schema, the validator, the retry logic, and the eval harness are the product
  • - structured outputs > free-form chat almost everywhere
  • - embed once, re-embed only when the source text actually changes
  • - measure on realistic inputs, not toy ones — bias toward end-to-end evals over unit tests of single calls

## skills

$ cat skills.json
{
"ml / ai" : [
"openai (gpt-4o, embeddings)" ,
"structured outputs / json mode" ,
"rag & vector search" ,
"prompt engineering & evals" ,
"pytorch" ,
"scikit-learn" ,
"huggingface transformers"
],
"languages" : [
"python" ,
"typescript / javascript" ,
"sql" ,
"go"
],
"data" : [
"mongodb atlas vector search" ,
"snowflake" ,
"airflow" ,
"dbt" ,
"pandas" ,
"spark"
],
"cloud / ops" : [
"aws (lambda, s3, rds, ec2)" ,
"docker" ,
"terraform" ,
"ci/cd"
],
"web" : [
"fastapi" ,
"react" ,
"astro" ,
"node.js" ,
"tailwind"
]
}

## what_im_looking_for

Open to full-time roles in applied ML, ML engineering, or ML-leaning full-stack. I take on select freelance projects too, especially around LLM features and search/retrieval systems. If something here lines up, I'd love to hear from you.

~  $  _
$ ls projects/ $ mail anthony →