Vicki Boykis Portfolio
Founding machine learning engineer at a startup working on recsys/personalization/information retrieval
Technical blog spanning 2015–2026 by machine learning engineer Vicki Boykis, documenting shipped side projects like Viberary (a vibe-based book search engine, since shut down and in maintenance mode) and posts working through embeddings and vector search, including "Querying 3 billion vectors," an experiment sparked by a Jeff Dean reply.
Location
Location: classified. Send a pigeon.
Companies
Skills
Notable
- Keynotes at PyCon Italia 2024, PyData Amsterdam 2023 and AMLC Charlottesville 2026
- Technical editor of O'Reilly's Building Recommendation Systems
- what_are_embeddings deep dive: 1.1k GitHub stars (paper DOI 10.5281/zenodo.8015029)
- Ran the Normconf ML/data conference
What makes it work
A breakdown of the choices that make this portfolio stand out.
A named side project with a public retrospective
Viberary isn't just mentioned once at launch — the January 2024 "Retro on Viberary" explains why the app was shut down (infra it would need next, dependency upkeep, $100+ a month in running costs) and lists what was learned, which is far more instructive than a typical "I built X" announcement post.
Concrete, quantified engineering problems as post titles
"Querying 3 billion vectors" and "How big are our embeddings now and why?" name the actual scale and question being solved, giving readers immediate context for the technical depth ahead rather than a vague thematic title.
A blog that documents production tradeoffs, not just successes
Writing about local model execution and infrastructure choices in practical, cost-and-tradeoff terms rather than purely celebratory terms reflects a working engineer's actual day-to-day decisions.
What Data Engineers can take from this
Specific, actionable tips to apply to your own portfolio — no generic advice.
If you ship a side project, write a retrospective months later on how it actually performed — that's more valuable to readers than the launch post alone.
Put the concrete scale or specific question in your post title ("querying 3 billion vectors") instead of a generic theme — it signals real depth immediately.
Write honestly about tradeoffs and things that didn't work as expected, not only clean success stories — it reads as more credible.
Projects
The post titles do the pitching here: "Querying 3 billion vectors," "How big are our embeddings now and why?" Vicki Boykis ships side projects the same way, including a public retro on shutting one down, which is rarer than the launch post.
A semantic search engine that recommends books by vibe ("atmospheric, female lead, worldbuilding, funny") instead of genre or title. It is a two-tower retrieval setup using a pretrained asymmetric MSMarco Sentence Transformers model, with training data prepared in DuckDB. Corpus embeddings were learned on AWS P3 instances, the model was converted to ONNX for low-latency inference, and vectors were served from Redis Search's HNSW index behind a Flask/Gunicorn API.
A full production path for an ML feature — data prep, embedding generation, ANN serving, load balancer, Nginx — built and operated by one person, plus a public account of why it was switched off.
ONNX conversion for inference latency; BERT-class model instead of an LLM ("Sometimes you don't need large language models")
Deployed on a single DigitalOcean droplet behind a load balancer
The retro calls out that evaluating unsupervised ranking is still unsolved
What are embeddings
A long-form deep dive into embeddings starting from fundamentals, published both as a paper ("Embeddings Approaches", DOI 10.5281/zenodo.8015029) and as a GitHub repo of notebooks. Described on the site as the groundwork that Viberary later put into production.
Explaining a core ML primitive from first principles, with a citable DOI, is a stronger signal of depth than using an embeddings API.
Published with a Zenodo DOI
Pairs prose with runnable notebooks
Querying 3 billion vectors
An attempt to reproduce the map-reduce approach Jeff Dean alluded to for querying 3 billion 768-dimensional vectors. It starts from a naive dot-product loop over 3,000 vectors, which took about 2 seconds on an M2 MacBook. Vectorizing the loop and casting to float32 brought 3 million vectors down to about 13 seconds, and the post extrapolates that to roughly 3,216 minutes for 3 billion.
It shows the working-out behind a system-design interview answer: measure the naive version, optimize it, do the memory maths, then question the requirements.
Benchmarks each optimization step with logged timings
Ends by interrogating the problem statement rather than claiming a solution
More Data Engineer portfolios
Ready to build your portfolio?
Follow the Data Engineer roadmap — skills, projects, and timeline to get hired.