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.
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 crew of coding agents on a Mac you own — a front desk that takes the work, a loop that breaks an approved plan into tasks, and a worker per task in its own session.
Stop prompting Claude directly. Reception takes the work, the loop plans it against an approved RFC, and every worker is a tmux session you can attach to mid-task and take over by typing. Approval happens in Linear; the factory reports what needs you. It ships no model and never sees your API keys — everything runs on the claude login you already hold.
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.
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.