A dead man's switch for solo devs, a small model with the AI tics trained out, and DeepSeek opens its inference stack

Welcome back to Builders Feed. Every day: what real builders shipped, why it matters, and what you could build next.

Shipped

1. Death Plan Switch: if a solo dev disappears, the right people are told

A one-person web studio owner in regional New South Wales built this after being hospitalised in 2012 and losing the business, because the data couldn't reach the backup developer in time. It emails you once a week. Miss it, and it reminds you at 24 and 48 hours, asks your next of kin "is everything all right?" at 72 hours, and only tells your other contacts (each with a message you wrote) at 144 hours, or straight away if your next of kin says something is wrong. Opening a link never counts as a check-in, because email scanners open links. It stores no passwords or documents, just where things are and who to call. PHP 8.1, no database, MIT licensed, written with Claude Code and covered by 42 unit tests. 70 points on Show HN.
Why it matters: Every one-person business has a bus factor of one, and clients know it. The builder says clients call it a strong sales feature when they trust a solo dev with their hosting and accounts.
deathplanswitch.com · GitHub · HN discussion

2. qwen3-4b-deslop: a small open model with the AI writing tics trained out

A fine-tune of Qwen3-4B-Instruct, trained on a MacBook Air, that writes without em dashes, "not X, but Y", or lists of three, and needs no system prompt. On 40 prompts it never saw in training, the author's slop score dropped from 121 to 12.5 per 1,000 words, em dashes went from 20.5 per 1,000 words to zero, and it still got 30 of 30 factual questions right. The write-up is honest about the failures: an earlier version scored better but wrote opinion pieces with no opinion and invented a university study. It runs locally with Ollama as a 2.5 GB download. 148 points and 100+ comments on Show HN.
Why it matters: Readers now spot AI writing on sight. This shows the habits can be trained out of the weights with a few hundred examples, and the write-up doubles as a guide to checking a fine-tune for damage, not just for the score you aimed at.
Write-up · Model on Hugging Face · HN discussion

3. DeepInfer: DeepSeek open-sources the inference framework behind its production service

DeepSeek released the in-house framework that serves DeepSeek-V4.1-Flash, under Apache-2.0. It includes their optimized kernels, attention and expert parallelism, draft-token decoding, batch-invariant inference (the same results no matter what else is in the batch), deterministic sampling, and structured outputs. It targets NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. The README is clear that this release covers offline inference, and that a full service still needs your own KV cache system, prefill/decode split, and load balancing. The announcement passed 6,500 likes on X, and the repo passed 4,400 stars within hours.
Why it matters: Model weights get shared all the time, but the serving stack is usually what labs keep. Having a production engine to read lowers the cost of running big open models for everyone who isn't a lab.
GitHub · Announcement on X

4. ai-job-search v1.7.3: a job-search agent locks itself down

A Claude Code framework that scrapes job boards, scores fit, tailors a LaTeX CV and cover letter, and has a reviewer agent critique the draft. The author built it after their position was cut in late 2025 and says it carried them through 69 applications and 20 first interviews to a job as an AI engineer. This release came after a public review by an AI agent concluded the shipped config was unsafe to run locally: skills now get only the tools they need, the permission file blocks uploads and credential reads in every mode, the agents that read postings can't edit or write, and every posting is checked for hidden instructions before anything reads it. It picked up 3,800+ stars today on GitHub trending. MIT licensed.
Why it matters: Any agent that reads text from strangers (job posts, emails, tickets) can be steered by that text. This changelog is a practical checklist for least privilege in an agent setup.
Release notes · GitHub

5. KiviDB: a multi-threaded, Redis-compatible database in Rust, with vector search built in

It speaks the Redis protocol (RESP2 and RESP3), so existing clients and redis-cli work unchanged, and it adds full-text and vector search with Redis Stack-style FT.* commands. The keyspace is sharded across threads, and it uses io_uring on Linux. The README's benchmarks, run on AWS Graviton bare metal, report about 21 million pipelined GETs per second on 32 cores. The builder says it started as a way to learn Rust after hitting memory blowups and OOM kills with Dragonfly at work. The core engine is AGPL-3.0, and it covers roughly 98-99% of Redis 8.10 commands. 68 points on Show HN.
Why it matters: Since Redis changed its license, teams have been weighing Valkey, Dragonfly, and others. One more drop-in option with vectors built in is useful, as long as you run your own benchmarks first.
GitHub · kividb.io · HN discussion

6. Tiny Little Office: ten small office apps where the document lives in the link

A word processor, spreadsheet, slides, outliner, calculator, drawing tool, diagrams, animator, calendar, and trip planner, all storing the whole document in the page's URL. Copy the link and you've copied the file. There are no accounts, nothing is saved unless you choose to keep it on your device, and saved short links are encrypted with a key the server never sees. Agents can use it too, through an llms.txt, an npx tinylittleoffice package, and a Claude Code skill. MIT licensed; launched on Product Hunt yesterday.
Why it matters: For a quick budget split or a one-page plan, a link that is the document beats setting up sharing permissions. And an agent can hand you a finished spreadsheet as a single URL.
tinylittleoffice.com · GitHub · Product Hunt

7. Yara Reads: a phonics reader for a six-year-old, and your agent can add books

A parent built this when their daughter finished her phonics app and there was nothing to read next made only of words she could sound out. Tap a word and its sounds light up one at a time. A checker tests every story word by word against the sounds the child has learned, by pronunciation, not just spelling. Her 12-year-old brother wrote the first story. Through MCP, you can connect Claude, ChatGPT, or another agent to write stories inside your child's sounds or type in books you own, visible only to your account. Runs on Cloudflare Workers with D1 and R2; MIT licensed.
Why it matters: It's a neat example of giving an agent a strict validator to write against, so you can trust what it produces for a picky audience.
yarareads.finereli.com · GitHub · HN discussion

Build this

Idea 1: A continuity kit for one-person IT shops

Death Plan Switch handles the "notice I've gone quiet and tell the right people" part. Build the rest for consultants and small MSPs: a per-client continuity pack that lists which tenant, registrar, DNS host, backup system, and break-glass account each client depends on (where they are, not the secrets), who the backup engineer is, and what to do in the first 48 hours. Keep it in an encrypted document with a second copy held by a lawyer or partner, wire it to the weekly check-in, and give each client a one-page summary they can show their own board. Death Plan Switch's builder says clients already treat this as a reason to sign; package it so every solo consultant can offer the same.

Idea 2: An injection checkpoint for agents that read the outside world

ai-job-search had to learn the hard way that a job posting can carry instructions. The same is true of every email, ticket, and web page an agent reads. Build a small gateway that every inbound document passes through before an agent sees it: it flags hidden instructions, strips invisible text, and labels the content as data. Pair it with per-agent tool allowlists, so the agent reading untrusted text can't write, send, or upload anything. Ship it as an MCP proxy plus a short audit report, and sell the setup to small businesses that are wiring agents into their inboxes and helpdesks.

Builders Feed is curated daily. Got something you shipped? Reply or tag @buildersfeed on X.