ProtoSpeech
ProtoSpeech
ProtoSpeech is a research & design studio building tools that are honest about what they do and respect the people using them — programming languages, communication protocols, creative tools, and open hardware, all shared openly.
Each project exists because something that should be simple isn't, something that should be yours isn't, or something that should be transparent isn't.
The projects furthest along — fully interactive, no setup required.
Most languages describe what you want, then a compiler secretly rearranges it into something else. Fudgel refuses that bargain — three sub-languages, each with one job, enforced by syntax. A slow program is a visible design problem, not a mystery buried in a toolchain.
Most compressors treat every byte the same. AIM decomposes data into what's structured and what's noise, building a mathematical fingerprint before choosing how to proceed. It beats gzip by 25–63% on audio and video, and correctly identifies when gzip wins on text.
Every major chat platform watches your conversations and locks your community into their ecosystem. Chirm is a self-hosted alternative built for the reading group, the game server, the neighborhood — communities small enough to deserve privacy and too important to hand to a corporation.
Bring your own image, and the toolchain generates a 3D-printable relief roller in seconds. Print in flexible TPU, roll through ink, press to paper. It's not a plotter — it's a tiny printing press that produces output with the tactile weight of something that came off real equipment. The machine is open hardware you can build for free.
A single encrypted .emb file holding passwords, notes, cards, bookmarks, and contacts. AES-256-GCM throughout, master key derived at 600,000 PBKDF2 iterations. Every item gets its own independently derived key. Developer access is architecturally impossible — there is no server, no account, and nothing ever leaves your device.
Content bundles carrying their own parser and template as hash-addressed artifacts. The host calls render(data, el) — that is the entire contract. Content can be JSON, Markdown, binary, or a custom DSL; the envelope never inspects it. Parser fails? Raw text. Template fails? Default styling. The system degrades gracefully at every step. Security comes from the browser sandbox; hash-based addresses make blocklists permanent.
Resolution, transport, and discovery layers underneath communication and content.
Email requires an entire infrastructure stack to self-host. Federation ties your identity to a platform. PING takes a different approach: one peer says "I have something for you," the other collects when ready. No accounts, no intermediaries. The sender can't tell if you ignored them or were just offline — and that ambiguity is the architecture.
Named for the Irish word for wind. Four message types — WANT, HAVE, CARRY, RESOLVED — propagate over any transport until a request is fulfilled or expires. Every node that touches a WANT takes responsibility for re-asking on the next wakeup. There is no failure state, only an unfulfilled want that can be re-issued.
A protocol where cryptographic identity replaces locational identity. The hash of a piece of content becomes its URL slug on every node that carries it — resolution cascades from local cache to local network to known nodes to origin. No new infrastructure, no protocol adoption required. If the hash is the path, any HTTP server already participates.
A video distribution layer built on three modes: static origin, WebTorrent swarm (every viewer seeds chunks to concurrent viewers for the duration of their session), and Deixis chunk streaming (content-addressed pieces resolvable by any node in the network). Origin bandwidth approaches zero for popular content. PING delivers direct follower notifications — no feed algorithm, no platform inbox, no middleman deciding who sees your work. A Phoric ecosystem project.
A decentralized domain directory seeded from TLD registry zone files — no crawl required for initial inventory, new registrations appear automatically from zone file diffs. An agent makes a minimal pass on each live domain to generate a structured record: title, description, headings. Because the metadata is agent-generated rather than site-owner-written, SEO is structurally irrelevant. Queries run as entirely local computation — nothing leaves your device. The index is a replicable methodology, not a proprietary database, so it cannot be captured or killed.
Systems and platforms that run on your own machine, under your own terms.
A local-first visual workspace where content lives in one place and gets referenced wherever it's relevant — never copied, never out of sync. Bounded pages force the organizational decisions that infinite canvases let you defer forever. Your entire workspace is a single portable file on your machine.
Modern operating systems are archaeological sites. ExactOS takes the desktop back to what it actually requires: draw pixels, handle input, track focus, launch things, manage clipboard, start at login. Each component does one thing and stops. No display manager, no D-Bus broker, no settings daemon, no system tray — under 15 processes at idle before any user app launches.
A self-hosted publishing platform with no platform at its centre. Your blog lives on your machine, signs every post with your local keypair, and distributes video via WebTorrent — no CDN bill that scales with virality. RSS is native architecture, not an export format. Reblogging carries the full signed provenance chain. Nobody can deplatform you because there is no platform holding your account.
A license composer that builds from zero permissions upward — you add only what you want to permit, so the final document contains exactly what you considered. Output is plain English that courts already recognize as legally meaningful. AI training clauses, commercial use, modification rights, and attribution conditions are first-class primitives, not afterthoughts. Each license gets a cryptographic timestamp, permanent public URL, and machine-readable output so platforms and AI pipelines can detect your intent automatically — without anyone reading a README.
Earlier-stage work and open questions, still taking shape.
A corpus toolkit that ingests ebooks, web pages, and documents — storing them as flat sequences of references into a shared token vocabulary. Fully lossless, trivially renderable, radically smaller than its source. Near-Shannon-limit addressing via a 255-symbol alphabet, O(1) lookup regardless of corpus size, and rich semantic metadata as a natural byproduct of import. Eight purpose-built databases, each with one job.
CRTs didn't lose to LCDs on image quality — they lost on form factor. Retrace replaces the electron gun with a scanning laser and MEMS mirrors, driving a flat electroluminescent phosphor panel. Real scanlines, real persistence, real bloom — not shader approximations, but material properties. No lead glass, no 25kV flyback, no vintage countdown clock. Solid-state, serviceable, manufacturable.
Four components — Corpus (complete, indexed by provenance), Engine (pure token-frequency mathematics, no semantics), Logic (a coherence gate that only rewards what can be grounded in the corpus), Identity (accumulated in artifacts on disk, not weights) — with no hidden state, no training procedure, no optimization target shaped by human feedback. The Engine generates prompts by seeking the sparsest regions of a frequency map: questions from genuine gaps, not learned associations. When something unexpected emerges from a system this minimal, the audit trail either explains it or it doesn't. Both outcomes are informative.
Animation, illustration, and whatever else escapes the lab.
Character animation, motion graphics, illustration, and graphic design — the visual and design work behind the projects, spanning several years across animation, game dev, and brand campaigns.
Visit ↗Standard doesn't mean optimal. It usually means first — or familiar enough that nobody reexamined it.
Every programming language, every protocol, every file format carries assumptions baked in decades ago — about how data should move, how code should be structured, how people should think. Most of those assumptions aren't wrong because they're malicious. They're wrong because they were never revisited. The tools got standardized, and the constraints they carried became invisible.
ProtoSpeech builds languages, protocols, creative tools, and open hardware that start by questioning those inherited defaults. The projects are different in scope and medium, but share a pattern: no black boxes, no lock-in, no intermediaries where none are needed — and a willingness to throw out the established approach when there's a better one hiding underneath it.
All work is shared openly under permissive or copyleft licenses.