r & d
The consulting practice funds focused infrastructure R&D, shipped under the hev brand. The client work keeps the R&D honest; the R&D keeps me sharp for the client work. Why these problems ->
The products. Each has its own site and its own lifecycle — this is the status board, not the pitch.
Voice-first assistants built around your own tools and sources.
A small framework for assistants with their own persona, tools, and visual identity. Voice, text, cards, and server-side tool policy are shared across dots for work, personal life, and travel.
A transparent, turbopuffer-shaped proxy that makes an existing vector store better without changing client code.
Your application keeps speaking the wire it already speaks; the gateway sits in the middle and adds retrieval. Automatic query routing reads the shape of a query and picks vector, lexical, or hybrid with RRF fusion per request. Typo-tolerant surfacing returns near-misses with a badge saying why. Query by id, radius scans, and point-in-time search round it out.
Your coding-agent transcripts, searchable, in infrastructure you control.
A capture daemon reads the session files your coding agents already write and indexes every prompt, reply and tool call into a turbopuffer namespace in your own account, with hybrid search running in hev layer. A local dashboard reads a trace back the way it ran, shows where the time went, and what it cost. hev up starts all of it on a Mac with Docker, and installs a skill so Claude Code and Codex can search your traces themselves. Kit hosts nothing and keeps no copy.
A ⌘K search overlay for docs readers, and the same docs as a markdown tree an agent can tree, cat, and grep.
hev ask distills a docs site into a committed ask digest — markdown a coding agent reads directly, and the index the overlay answers readers from. A Go core with an Astro adapter, published as the npm package @hevmind/ask. Free and open source. Astro is wired end to end; other framework adapters are designed, not shipped.
A YAML-first eval framework for any LLM — compare models, mix string matching with semantic and LLM-judge checks, run it from the command line.
Built for the tightest feedback loop I could get on a prompt change: did this get better or worse, and on which models? The CLI ships on npm as vibecheck-cli and the DSL still runs. The hosted preview is winding down — what it taught me about eval design went into how I run evals on client engagements instead.
Not a hev product and not a client engagement — a whole site, one operator directing a fleet of agents, kept deliberately outside the hev design system because the range is the point.
A travel publication, built whole — destination guides, illustrated art direction, search, accounts. Next up: an agentic trip app builder.
One operator directing a fleet of agents, and the brief is a whole product rather than a feature. It is deliberately not in the hev design system: when the brand calls for loud, ship loud.
-> travels-with-charlie.comI build source-available tools for benchmarking, comparing and evaluating database infrastructure, to help developers, IT leaders and entrepreneurs make decisions about technology. No content here is ever sponsored — but I do work with vendors to help them isolate and improve the performance of their products. Get in touch if that's interesting to you.
Opening one takes a free account — name, title, and what you're working on. That last field is how I pick what to build next. Sign up.
Build your own plots from every accepted run in the reranker benchmark: quality, price and latency for 50 systems on SciFact, NFCorpus and FiQA.
Hosted rerankers, open-weight rerankers, decision models used as rerankers, and listwise LLMs, all reranking the same BM25 top 30. Pick the axes and the systems, start from the charts in the posts, and download the plot or the data. Every number links to its receipt; costs are list-price estimates, not invoices, and latency comes from different hardware, which the page says next to each point.
The data behind "Hev meets Jev", "Jev has company" and "OpenAI makes four".
What a vector index actually costs, month 1 to month 36 — starting from how much data you have, not how many vectors you think you'll store.
Vendor calculators quote a month you will never have: they ask for a vector count nobody knows and price the steady state. The bill that surprises people is writes — how much data your index actually creates, once, and then keeps forever, times every re-index. Five inputs, 36 months, list prices, every constant shown. Queries are deliberately not modelled: they scale with traffic rather than corpus size, and they are not what surprises anyone on the invoice.
Built for a Maven Lightning Lesson with Doug Turnbull, August 2026.
Why these problems and not others: I only take on R&D that clears three bars.
All three, not two of three. What that leaves runs in two halves. Development is the products above. Research is the tools I build to answer a question and publish so you can check the answer.