Hyphen
Live- Streaming RAG
- Multimodal
- Cost engineering
- Reliability
A customer-support intelligence platform that turns conversations into structured, actionable signal.
- 2 months
- To first delivery
- $1M / 9mo
- Competing quote
- 117k
- Lines of code
AI Engineering · Seattle, WA
Hugo
Pretorius
Retrieval, agents and evaluation — shipped, not prototyped.
The decade behind it: Flutter and mobile engineering — lead and manager at SKIMS, instrument clusters across Toyota and Lexus — the same shipping bar, now pointed at AI systems.
I build production AI systems end to end — retrieval, agents, evaluation and the infrastructure around them. Over the last twelve months that has meant six systems built solo, one of them a support-intelligence platform delivered in two months where the competing agency quote was nine months and $1M. Before this I led mobile engineering: a consumer app rated 4.9 across 28,000 reviews, and instrument cluster software shipping across Toyota and Lexus vehicles.
Selected work
All of it →Yolk Studio
- Agents & tool use
- Human-in-the-loop
- NL→SQL
- Multi-tenant
A multi-tenant commerce SaaS with a tool-using agent and a natural-language analytics engine built in.
- 3,076
- Commits, solo
- 11 months
- Sustained
- 15
- Agent tools
Memo
- RAG with citations
- pgvector / HNSW
- RLS-scoped retrieval
- Whisper
Queryable organisational memory, built to remove cross-team silos, with citation-backed answers.
- 23 days
- Build time
- 59k
- Lines of code
- 15
- ADRs
svelte-a2ui
Live- Generative UI
- Untrusted-input hardening
- Protocol design
- Open source
The Svelte 5 renderer for A2UI — agent-described interfaces rendered from data, never from generated code.
- 7 days
- Idea to npm
- 83
- Protocol tests
- 18
- Catalog components
auri
Live- Schema design for LLMs
- Evals in CI
- Structured output
- Open source
Agent-facing component catalogs for A2UI, designed by putting the schema in front of models before building anything behind it.
- 5
- Gate findings logged
- 2
- Model families in CI
- 35
- Commits
Writing
Arcanum →Embeddings · 7 min
How do I tune a threshold when I only logged the matches?
A similarity score was recorded every time two support cases were grouped and discarded every time they were not, so every number available for tuning was, by construction, above the threshold. The three near-misses that moved it were invisible until one column kept them.
Retrieval · 3 min
How do I get a model to answer from my own documents?
A model has never seen your documents, so asking it about them produces fluent, confident invention. The fix is not to teach it — it is to hand it the right passages at the moment you ask, which is the whole idea behind retrieval.
Evaluation · 4 min
How do you evaluate a retriever without labels?
Everyone agrees retrieval should be measured, and almost nobody has a labelled set to measure against. Twenty questions written by hand beat two thousand generated ones — a synthetic question is written from the passage it is meant to find, and shares its vocabulary in a way a real one never does.