public digest · 5 picks
Real-time offline translation, Elixir caching, and a single-binary Postgres explorer
A theme emerged in this week's batch without us planning it: less cloud dependency, more control. The standout example is RTranslator running Whisper and NLLB entirely on an Android phone, but pgweb shipping as a single Go binary and Nebulex keeping your cache adapter out of vendor lock-in follow the same instinct.
The other thread is 'useful but honest about its state.' Nextacular's custom domain support is real and saves days of work, but the deps haven't moved since 2022 — that matters when you're building billing on top of it. hardhat-deploy v2 makes EVM deployments dramatically cleaner, but it requires Hardhat v3 which is still in alpha. We're flagging these things upfront so you can make the call with full information.
// pick 1 of 5
RTranslator is an Android app that does real-time speech translation entirely on-device using Meta's NLLB for translation and OpenAI's Whisper for speech recognition. The main mode connects two phones over Bluetooth LE so each person hears the other speaking their own language through headphones. It's for anyone who needs offline, private translation — travelers, field workers, situations where cloud APIs aren't acceptable.
The core engineering work here is serious. Getting NLLB and Whisper running simultaneously on a phone with partial int8 quantization, KV cache, and split model loading — and hitting 1.3GB RAM with 4x speedup over a naive ONNX conversion — is not a weekend project. The privacy story is also genuinely real: no account, no server, no data leaving the device. The Bluetooth LE peer-to-peer mode where each person hears the other in their own language through headphones is the kind of thing that sounds like a demo until you see it working.
What to know going in: the 6GB RAM floor is a hard cut that excludes most of the world's Android devices, which is an uncomfortable mismatch given the use cases (field work, travel, situations without reliable internet). The known Bluetooth dropping bug has no fix timeline listed, and for the primary use case — two people mid-conversation — a dropped connection is a serious failure. TTS quality is entirely dependent on whatever system TTS is installed, which varies wildly. The 1.2GB model download pulls from GitHub on first launch, which fails silently in regions with GitHub connectivity issues.
View on GitHub →
// pick 2 of 5
Nebulex is a caching abstraction layer for Elixir, modeled after Ecto's adapter pattern. You define a cache module backed by a swappable adapter (local ETS, Redis, Memcached, etc.) and get a consistent API plus declarative decorator-based caching across all of them. Aimed at Phoenix/Ecto developers who want Spring Cache-style annotations without vendor lock-in.
The decorator approach is the best part: `@decorate cacheable(cache: MyCache, key: {User, id})` sits cleanly at the function head and keeps cache logic out of your business logic without feeling like a hack. The adapter pattern is properly thought through — adapters live in separate packages with their own release cycles, the core stays thin, and swapping from local ETS to Redis doesn't touch your application code. The `references:` option for secondary-key lookups pointing at primary-key entries is a subtle detail that saves you from the classic cache consistency headache where a username lookup and an ID lookup diverge.
What to know going in: the core package ships only a `Nil` adapter — you pull in `nebulex_local` or a third-party package for anything real, which adds a setup step that trips up first-timers. The distributed topologies (partitioned, replicated) live in separate repos with independent release cycles, so 'distributed caching toolkit' is more of a description of the architecture than a thing you get out of the box. The generational GC eviction model (rotating 'generations' every N hours rather than LRU) reliably surprises people coming from other caching libraries — memory usage effectively doubles at generation boundaries, and it's worth reading the docs on this before you size your cache.
View on GitHub →
// pick 3 of 5
hardhat-deploy is a Hardhat plugin that manages smart contract deployments across EVM chains, tracking addresses by name instead of index and enabling deterministic, replicable deploys. v2 is a full rewrite built on top of rocketh, a framework-agnostic deployment layer. Aimed at Solidity developers who need repeatable multi-chain deployments with proxy and Diamond pattern support.
Named accounts replacing magic array indices is a small change that meaningfully improves every deployment script and test that touches account management — `deployer` reads better than `accounts[0]` and doesn't silently break when your account order changes. The Diamond (EIP-2535) support is the more substantial win: declaring your target facet state and having the plugin generate the `diamondCut` call eliminates a class of error that's easy to make and hard to debug. The pending transaction retry logic is genuine production tooling, not a demo.
What to know going in: v2 requires Hardhat v3, which is still in alpha — if you're on a stable production project, you're on the v1 branch, and v1 is clearly not getting new features. The rocketh layer adds real conceptual surface area: you need to understand hardhat-deploy, rocketh, and the individual `@rocketh/*` packages before you can trace a deployment failure end-to-end. The Hot Contract Replacement feature works cleanly within the expected proxy conventions and quietly stops working outside them — worth understanding before you build a deploy workflow around it.
View on GitHub →
// pick 4 of 5
Nextacular is a Next.js SaaS starter kit with multi-tenancy, Stripe billing, custom domains, and team/workspace management baked in. It targets developers who want to skip the plumbing and ship an MVP faster. The stack is conventional and well-understood: Next.js, Prisma, Tailwind, NextAuth.
Custom domain support per workspace is the single most valuable piece here, and it's actually implemented rather than stubbed. Getting custom domain routing right — DNS verification, certificate provisioning, per-workspace routing — takes longer than most of the other SaaS plumbing combined, and having a working reference to read is worth a lot. The Prisma schema gives you a real data model as a starting point, Stripe webhook handling is present and wired up, and NextAuth with magic links works end-to-end.
What to know going in: the dependencies are multiple major versions behind across the board — Next.js 13.1, Prisma 4.4, Stripe SDK 10 — and the last migration is from 2022. That's not a minor warning; it means you're inheriting the upgrade work as part of adopting this. There are no tests anywhere in the codebase, which is a real problem when billing and auth flows are the core of what you're building on. The multi-tenancy subdomain routing is tied to Vercel's platform, so self-hosting requires reworking that layer. Treat this as a well-structured reference rather than a drop-in foundation you ship without changes.
View on GitHub →
// pick 5 of 5
pgweb is a single-binary, web-based PostgreSQL explorer written in Go. Zero runtime dependencies — download one file, point it at a database, and you have a browser UI for browsing tables, running queries, and exporting data. Aimed at developers who want something lighter than DBeaver or TablePlus without installing Electron.
A single binary you can drop on a remote server and connect to through SSH tunneling — with the SSH tunnel support built in rather than requiring a separate tunnel process — is genuinely useful in the situations where you need it most. No Electron, no Node, no Python runtime. Native multi-session support lets you hold connections to multiple databases at once without opening multiple terminals. The CockroachDB handling (separate statement files in `pkg/statements`) suggests the wire-compatibility story was thought through carefully.
What to know going in: the frontend is Bootstrap 3, jQuery, and inline Ace editor, and it reads that way — this is a functional tool that hasn't been modernized in years, not something you'd demo to a non-technical stakeholder. It's a pure data browser; no pgvector, no logical replication slots, none of the Postgres 14+ features. Query history lives in memory only and disappears on restart, which limits its usefulness as a daily driver. The connect-backend auth integration has documentation in the form of two code examples in a config folder — usable if you read the source, but not something you'd trust in a shared team deployment without doing that homework.
View on GitHub →
That's the five for this week. If you want these in your inbox every week without remembering to check, there's an email signup at the bottom of the page — no other emails, just the digest.
Get this in your inbox →