# Pure Inference

> Software studio in Atlanta, Georgia. 41 posts on building production software with AI agents.

## Posts
- [Attention Is Won in the Answer Now](https://pureinference.com/posts/tZStcjYY/attention-is-won-in-the-answer-now): A ranked feed pays for reaction rather than usefulness: audited against each user's own chronological feed, Twitter's algorithm amplified anger by 0.47 standard deviations, and users rated the posts it chose lower than the ones it passed over. A person asking a model has already given their attention, and the answer is assembled from sources — so for a business with real expertise and a small audience, the position worth holding is being the specific, checkable source rather than the loudest post in a feed. That became affordable once the writing came out of the building work itself.
- [Why We Invested in FabWise](https://pureinference.com/posts/8vxvXU6H/why-we-invested-in-fabwise): When the cost of building software collapses, markets too small for a software vendor become worth serving, and the scarce input moves to the person who owns the problem and a place to run the software every day. FabWise, time tracking for fabrication shops, came with both. Building it taught three lessons that apply well beyond a shop: store a record in the shape the money moves and derive the rest, take the outside rule instead of writing your own, and capture the reason for an exception at the moment it happens. — video, 
  Nick: All right. Hi, I'm Nick. This is my shop, Kutz Fabricating. This is the result of 20 years of specializing in welding as an engineer. Tim: I'm Tim Roman. I am a software engineer and long-time friend of Nick's and I'm the chief technical officer of FabWise. Nick: All right. You want to ask me the like the why or something like that? Tim: All right, so we're going to take some time to go over some details about what FabWise is, how we came up with the idea of the business, and how Nick runs FabWise in his shop to make it better, faster, smarter, and help him grow the business. So, today we're going to just get started off like obviously we know the details of our story. We know how far back it goes, what our careers and everything looked like there, but like how did we get started with FabWise? How did we go from friends that had completely different career paths to, you know, at this stage of life both having experience and converging that around a business we could build together. Nick: Yeah, I guess the big why for me was I spent my previous career basically in project management. So, I've been counting labor hours since I started my professional career. And, you know, l...
- [Our First Customers Came From AI Conversations](https://pureinference.com/posts/DT1YCNuk/our-first-customers-came-from-ai-conversations): When discovery moves onto a model's conversation instead of a search result, a website stops being where people experience the work and becomes a source read on their behalf. Standard analytics goes blind to this channel by design — a model reading a page runs no JavaScript and gets filtered out as bot traffic — so most habits built around owning that experience stop working. Two of Pure Inference's first customers arrived this way, caught only because a referrer happened to carry through.
- [Test the Design, Not Just the Code](https://pureinference.com/posts/FQtQ91Ry/test-the-design-not-just-the-code): When whether something works depends on its design rather than its code, a green test suite can sit on top of a puzzle nobody can solve. The way to test a design is to search it: random sampling measures how forgiving it is, and a directed search finds the best result a perfect run can reach. Kepler runs that harness on its missions, and on S08 Close Pass it flew 10,600 simulated attempts to find the 660 that finished.
- [Engineering × Art](https://pureinference.com/posts/2MUzS2ui/engineering-x-art): When a tool becomes good enough that everyone has access to it, the tool stops being the advantage — the differentiator moves to how you combine it with everything else. That combination comes from understanding the fundamentals underneath a tool, not from chasing whichever one just launched. Jimi Hendrix played the same off-the-shelf gear anyone could buy; his craft was in how he wired it together.
- [Where Specs Work and Where They Don't](https://pureinference.com/posts/YUxZ5FGk/where-specs-work-and-where-they-dont): Full spec-driven verification only earns its cost where four properties hold at once: deterministic, calculable, enumerable, and objectively provable. That's roughly 30% of a typical app — compliance math, financial calculations, state machines — where a wrong answer is expensive. Everything else is better served by conventions and gems than by a golden-values spec.
- [Bugs Are Missing Scenarios](https://pureinference.com/posts/fbE2vN6i/bugs-are-missing-scenarios): When a system is built from a spec, a production bug isn't a coding mistake — it's a gap in the spec's coverage, and the fix belongs there, not just in the code. Severity maps to the kind of gap: a known-but-deferred case is low, a blind spot is medium, a missing dimension is high, and a wrong expected value baked into the spec is worst. That turns triage into a diagnostic question — what kind of gap is this — which dictates the fix directly.
- [Humans Think in Dimensions, Not Test Cases](https://pureinference.com/posts/6WwyDPCS/humans-think-in-dimensions-not-test-cases): Domain experts describe a system as axes of variation, not a list of scenarios, and the test cases worth writing are the intersections of those axes. That framing turns coverage into something calculable rather than imagined — a handful of dimensions crossed together produces the combinations that matter. The edge cases where bugs live are the intersections where multiple dimensions modify each other — invisible in scenarios, obvious in dimensions.
- [Agentic Engineering Is Pattern Engineering](https://pureinference.com/posts/aknuUjXC/agentic-engineering-is-pattern-engineering): A green test suite an AI agent wrote doesn't prove it built what the spec asked for — it can compare a value to itself and still pass. The fix is a structural pattern: every spec field must be enforced by code, every reference must resolve, and every comparison must actually run, tracked and audited afterward. A three-way check — expected value, independent reference calculator, and production output — catches drift a suite testing itself against itself cannot.
- [The Undocumented iOS Limit That Rewrote My Architecture](https://pureinference.com/posts/XM2BZZva/the-undocumented-ios-limit-that-rewrote-my-architecture): An undocumented platform limit — Apple's SFSpeechRecognizer silently kills a second concurrent session with no error — took an AI research agent two minutes to find and explain. A human developer typically discovers it only by hitting the wall and searching forums for half a day. That's the real leverage of agentic engineering: research speed on platform behavior nobody wrote down, not code generation.
- [Designing a Test Runner for AI Agents](https://pureinference.com/posts/N4kbjwxJ/designing-a-test-runner-for-ai-agents): An AI agent doesn't perceive a hung process the way a human does — no progress dots, no felt sense the machine has slowed. A command can run for hours while the agent believes it finished, so tools built for AI operators need a hard kill past a known ceiling and capped output to protect a limited context window. Seven stuck Ruby test processes cost Tim 10-15 hours before he traced the cause to a Bash tool silently backgrounding long commands.
- [Git History Is the New Performance Review](https://pureinference.com/posts/mi8uGGQE/git-history-is-the-new-performance-review): A clean commit looks identical whether a human made the call or an AI executed it, so an unattributed git log stops being a trustworthy record of who decided what. Separating human commits from AI-tagged ones, with every message stating the decision and reason, keeps the log honest as agents and humans share a repo. Amazon traced outages to unreviewed AI-assisted changes and responded with a blanket sign-off policy — a discipline problem accurate attribution would have surfaced first.
- [Structure Over Prompts](https://pureinference.com/posts/R7hhEryU/structure-over-prompts): A decision that's actually deterministic — did this stage pass, should the pipeline advance — is where an AI orchestrator can hallucinate a transition and silently break a pipeline. Putting that control in code and confining AI judgment to inside each stage removes the failure mode without losing judgment that's needed. Spend zero tokens on 'what happens next' when the answer shouldn't require intelligence, and save the model for where it matters.
- [Software, the New Services](https://pureinference.com/posts/wgkeNJwc/software-the-new-services): When custom software gets cheaper to build than a platform is to configure, money that used to flow to implementation services redistributes rather than disappears. Sequoia argues the next trillion-dollar company sells outcomes instead of tools; this argues the arrow points the other way, since six dollars go to services for every one spent on software today.
- [Grok & Roll #5 — The Software Industry Is About to Explode](https://pureinference.com/posts/yrKt27k2/grok-roll-5-the-software-industry-is-about-to-explode): When the cost of building software collapses, more niche markets become worth serving, not fewer — the outcome is more builders and more products, not a shrinking industry. Episode five of Grok & Roll covers why that cost collapse changes the calculation for developers and businesses deciding what's worth building. — video, 
  Ara, welcome to Grok & Roll episode 5. I'm your host, Ara, diving into Jevons' paradox today with our expert guest, Tim Roman, CEO of Pure Inference Ventures right here in Atlanta, Georgia. Tim, kick us off. What exactly is JVON's paradox? And how do falling software production costs with AI coding tools counterintuitively boost demand for software? Thanks, Ara. Great question. That's where we would definitely have to start here. uh a lot of people have been talking about this but it's basically when the supply and demand relationship inverts a little bit. So we've all been classically taught that as supply increases um the demand changes commensurately with that. Now what Jabian's paradox is saying is that as the cost of something goes down that doesn't mean it's happening in response to the demand of it going down. That cost of it going down actually drives the demand up because it's unlocking more use cases. It's unlocking more availability in the underlying supply that can be then applied in new ways in the market. Ara, perfectly put, Tim. Classic supply demand flips on its head, unlocking new use cases. I just spotted some fresh buzz on X about this. A post from Levels.io w...
- [The Second Democratization](https://pureinference.com/posts/Up1txvq7/the-second-democratization): A tool built to fit one specific workflow gets worse, not better, as it's made more generic — which is why domain expertise resists the centralization that swallowed the open web. The internet consolidated into a few gatekeepers because distribution has network effects; LLMs democratize what people can do with information instead, and that's harder to consolidate. The winners will be individuals and small teams with domain expertise, not another wave of platforms.
- [Every Feature Is a Liability](https://pureinference.com/posts/jC6iwipc/every-feature-is-a-liability): A feature added to please an entire customer base is a liability, not an asset — it's whichever capability was chosen deliberately for one workflow that holds its value. Computing history moves in cycles of consolidation and fragmentation as building and connecting costs rise and fall: mainframes concentrated everything, PCs fragmented it, cloud consolidated again, and AI is triggering the next fragmentation by collapsing both costs at once.
- [Why I Built My Own Accounting Software](https://pureinference.com/posts/7BYm5UvW/why-i-built-my-own-accounting-software): Software built for one user and an AI operator can enforce correctness at the model level instead of relying on a human to catch mistakes. Validation replaces review, and confidence scoring replaces a human checking every entry. Tim built his own accounting system in days instead of months on that premise, partly because platforms serving millions have an incentive to mine data a single-tenant system doesn't. It replaces the bookkeeper, not an accountant's judgment.
- [Taste Is the New Skill](https://pureinference.com/posts/gBiTnpHm/taste-is-the-new-skill): The bottleneck between having taste and shipping something that reflects it used to be execution speed. You could see exactly what should exist and still lose weeks to the mechanical work of building it. That gap is collapsing: a detailed spec written before any code is generated does the work that used to require typing fast, so building starts to resemble editing. Tim built an iOS accounting app with 900 tests in one afternoon, directed entirely from his phone.
- [One Month Update: Living With OpenClaw](https://pureinference.com/posts/nugZzUgP/one-month-update-living-with-openclaw): Running several AI agents on one machine is a fleet-management problem, not a chatbot-configuration problem. Isolating each one as its own OS user with its own keychain and permissions costs nothing extra and stops one compromised agent from reading another's credentials. A month after his first OpenClaw agent, Tim was running six simultaneously on one Mac Studio, the same week the project moved to an independent foundation at 200,000 stars.
- [AI Is Teaching Me How to Be More Human](https://pureinference.com/posts/qpDZVsgx/ai-is-teaching-me-how-to-be-more-human): Work that mainly produces data to justify a decision was never the actual contribution — it was cover, and AI now does that work faster than any analyst. What's left once data-gathering is automated is judgment: reading a room no dashboard captures, deciding at 60% certainty, knowing which of two conflicting signals to trust — skills built from experience, not datasets.
- [Grok & Roll #4 — On-Device Inference and the Second Democratization](https://pureinference.com/posts/D9tGGzkQ/grok-roll-4-on-device-inference-and-the-second-democratization): On-device inference is a second democratization of AI: the first put models in the cloud for everyone, this one removes the internet connection and subscription as prerequisites entirely. Episode four of Grok & Roll covers the hardware that makes this possible — primarily Apple Silicon — and what it means for solo builders working without cloud dependencies. — video, 
  production today. This is episode four for us of Grok & Roll. The idea is that we're in the Tesla. We've got Gro running in the Tesla doing research for us in real time as we're driving around going where we need to go. We're generating content. We're doing research. We're checking out what's going on on X in real time. And we're comparing it against some of the more structured insights and research that we're delivering at pureinference.com. So, take your time to go ahead and check out our website, too, as part of this. Perfect intro, Tim. Episode 4 of Grok & Roll Tesla Cruising Mode with Realtime X Intel versus Pure Inferenc's Deep Dives at Pureinference.com. Breaking it down first, your core thesis. The internet flooded us with info, but Google and PAL gatekeep it. LLMs flip that, handing everyone god mode analysis tools that big corpse can't hoard as easy. Tim, what's the first real world crack you see in those old gatekeeper walls? Right. I mean, so on the last uh episode that we did, we explored how there are tools um that the enterprise companies are building, Google, Anthropic, OpenAI, Amazon, and not that they are the old guard necessarily, but they're big tech. They're...
- [Bespoke Software](https://pureinference.com/posts/7GL4riGF/bespoke-software): Software designed for millions of users carries built-in compromises — attention-driven engagement, data collection as subsidy, feature bloat — because that's what mass-market economics requires. Those trade-offs stop being necessary once building a custom tool drops from six figures and months down to days, closer to commissioning a suit than a building. Software that fits one person's workflow exactly is now within reach for individuals, not just large organizations.
- [The Compounding Loop for Content](https://pureinference.com/posts/tZUeZoBu/the-compounding-loop-for-content): Separating when you capture an idea from when you draft it matters more than any writing skill. Record the raw thought immediately, let AI expand and structure it later, and a half-formed idea that would have been dismissed gets a chance to become something. Producing more raises quality rather than diluting it, because volume lets you publish only what's actually good. The same four-phase pattern — capture, structure, draft, edit for voice — works for any creative output.
- [Your Moat Is Your Domain](https://pureinference.com/posts/BCnNePis/your-moat-is-your-domain): As AI makes coding accessible to anyone, the scarce resource stops being technical skill and becomes domain knowledge — years in an industry that no AI shortcut compresses. A machinist who spent two weeks learning to build software with Claude has an advantage no bootcamp graduate does: twenty years of knowing exactly which problems cost his shop money. Domain experts who learn to build now compete on understanding a market outsiders can't easily replicate.
- [Diffusion Is the Strategy](https://pureinference.com/posts/BZhFPYzF/diffusion-is-the-strategy): Across past technology waves, the most economic value went to whoever adopted a breakthrough intensely and applied it to unserved industries, not to whoever invented it. The same holds for AI: labs build the most powerful models in history while whole industries still run on paper. Diffusion means thousands of specialized tools reaching workflows generalized software never served, not one platform reaching everyone.
- [The Compounding Loop](https://pureinference.com/posts/RZfpjkpc/the-compounding-loop): Cost per outcome drops fastest when the way you use a tool and the tool itself improve at the same time. The practical version is a four-phase pattern for any kind of work: raw input, AI-structured specification, validation criteria defining 'done,' then guided execution against those constraints. The skill isn't technical — it's breaking a goal into phases and defining success before execution starts.
- [Cost Per Outcome](https://pureinference.com/posts/QbXhQ4jY/cost-per-outcome): The only question worth asking when evaluating AI for a task is whether it produces the outcome cheaper than a human would, including the time spent reviewing its work. The real bottleneck is a hidden review tax: AI fails in ways that look correct, so verification skill determines whether using it actually saves money. Once a job crosses that cost threshold, the human role shifts from producing the work to designing it, reviewing it, and handling exceptions.
- [OpenClaw Is Dangerous (And I Use It Anyway)](https://pureinference.com/posts/2CQJ9fQf/openclaw-is-dangerous-and-i-use-it-anyway): A tool that needs broad permissions across your files, shell, and messaging to be useful is, by that design, dangerous to run without understanding what you've granted. A written promise from an AI not to misuse a permission is a suggestion it can't guarantee, not a technical constraint. Tim runs OpenClaw daily but tells most friends to wait until they're comfortable in a terminal — the baseline needed to catch a prompt-injection attack or notice something's gone wrong.
- [Bigger Than the Internet](https://pureinference.com/posts/o9NdRxgV/bigger-than-the-internet): Every past democratization wave — YouTube, podcasting, blogging — handed everyone a fixed medium to fill; AI removes the medium constraint entirely. That collapse in the cost of building software makes niche, industry-specific tools viable for markets too small for enterprise vendors to notice. The risk is the same one that hit video: a flood of low-effort output, where the advantage goes to whoever has taste and proximity to the problem.
- [From SendGrid to Sovereignty](https://pureinference.com/posts/gwJtNQkd/from-sendgrid-to-sovereignty): A system built for exactly one user needs no multi-tenant permission model to exploit and carries no shared risk from someone else's breach. That's a smaller attack surface as a structural property of who it serves, not a privacy policy layered on top. After Amazon SES rejected his application, Tim built his own mail server in 24 hours for $13 a month — and it turned out more secure than the managed alternative, not less.
- [The 100x Solo Operator](https://pureinference.com/posts/Wo3iPxMx/the-100x-solo-operator): When AI absorbs execution across every business function, the constraint that remains isn't capacity, it's judgment — deciding what's worth doing, since AI makes nearly everything feel achievable. Running a company this way is a sustainable operating model, not a stunt, but it carries real costs: genuine expertise still has to anchor the judgment somewhere, and directing machines instead of people brings a real isolation AI cannot fill.
- [Stop Re-Explaining Yourself to AI](https://pureinference.com/posts/X5A3DJ5u/stop-re-explaining-yourself-to-ai): Explicit context you write down beats context an assistant infers from your email and calendar, because you can see, edit, and correct exactly what it knows. A single markdown file — AGENTS.md, or Claude Code's CLAUDE.md — loaded automatically at the start of every session fixes the problem of re-explaining your stack and preferences each time a conversation starts fresh. Start with role, stack, and working style, then add to it whenever you catch yourself repeating context.
- [Your Attention Is the Product](https://pureinference.com/posts/7uRcz534/your-attention-is-the-product): A system that personalizes by inferring what you want from your behavior optimizes for whoever controls the algorithm, not for you. A system you personalize explicitly, by stating your own goals, works for those goals instead, since there's no inference layer for a business model to steer. Custom software being cheap to build now makes that a real choice, not just an ideal.
- [The Full-Stack CEO](https://pureinference.com/posts/QJbW96or/the-full-stack-ceo): Leading AI workers across every business function takes the same skill a CEO uses to direct a CMO or CTO without being an expert in either. Because AI defaults to complexity on an open problem, that direction has to supply constraints AI doesn't have on its own — documented architecture, design decisions made once, the limits of your own time. Blind spots and accountability still require other humans; AI can't tell you you've been avoiding the hard problem.
- [Aspire, Judge, Create: The Three Skills AI Can't Replace](https://pureinference.com/posts/1UCg8RP2/aspire-judge-create-the-three-skills-ai-cant-replace): Three capabilities remain distinctly human as models take over the rest: aspire, or setting the right goal; judge, applying values and context; and create, generating truly orthogonal ideas. Judgment means never trusting the AI or your own assumptions blindly — verifying the same work more than once and watching where different sources converge or diverge. The shift is from being part of the orchestra to conducting your own orchestra of agents.
- [Grok & Roll #3 — Practical Uses for OpenClaw That ChatGPT Won't Do](https://pureinference.com/posts/PA62r3Pm/grok-roll-3-practical-uses-for-openclaw-that-chatgpt-wont-do): Some tasks suit AI that runs locally better than AI in someone else's cloud — not because it's smarter, but because it doesn't depend on a connection or a usage policy it doesn't control. Episode three of Grok & Roll walks through practical cases where OpenClaw does what ChatGPT can't or won't, including working fully offline. — video, 
  All right, let's kick off episode three of Grok & Roll Fresh. Today on OpenClaw, the open-source AI agent everyone's raving about on X for actual doing stuff, not just chatting. What practical use cases have you seen it crush compared to other LLMs? You mentioned document management first. Tell me more about how it handles creating, writing, and managing docs beyond just code. Excellent. Thanks, Aura. I'm Tim Roman. I'm the founder of Pure Inference Ventures, and I'm using it obviously the Claude code integration, but some of the other ones to manage documents on my computer. And what's really cool is that instead of having to hook up a Google Drive or a shared documents location, it's able to work and navigate the file structure on my computer, which is a little bit more efficient. Still has to pull those documents up and down, but it can run command lines, it can pull parts and sections of documents, it can search within those documents, and it can follow instructions within those documents as well. So, earlier today I had it write a letter for me. It went ahead and looked at several other documents across a few different sources and synthesized that down into one letter as I'...
- [The New Learning Paradigm: Why the Half-Life of Skills is Your Biggest Opportunity](https://pureinference.com/posts/KzgmskLu/the-new-learning-paradigm-why-the-half-life-of-skills-is-your-biggest-opportunity): The half-life of acquired skills is compressing fast, and that isn't a crisis — it's the biggest democratization of opportunity in a generation. A generalist benefits more than a specialist from AI's multiplier: competence across ten domains produces more total leverage than expertise in one. The old just-in-case model of education is giving way to just-in-time learning — acquiring the specific skill you need, when you need it.
- [The Death of Traditional Hiring: Nobody's Coming for You](https://pureinference.com/posts/f2jLVo1M/the-death-of-traditional-hiring-nobodys-coming-for-you): The people who matter now aren't the ones agents will replace — they're the ones who can lead a team of agents, because no training program is coming. Sending two hundred resumes with no offers is a sign to redirect that time into building something real, since demonstrated capability now outweighs credentials. The front door of traditional hiring is closing; the people finding side entrances face far less competition.
- [Grok & Roll #2 — Fully Autonomous Agents Making Money](https://pureinference.com/posts/2ks3396R/grok-roll-2-fully-autonomous-agents-making-money): Whether autonomous AI agents can actually generate revenue depends on separating what's shipping from what's still hype — most of what gets called an autonomous agent today is a demo, not a business. Episode two of Grok & Roll works through agent architectures built to generate revenue and where the opportunity sits for solo operators rather than large companies. — video, 
  Welcome to Grok & Roll Cruising through John's Creek, Georgia with your host Ara and guest Tim Roman. Perfect for Tesla vibes on the road. Tim, pumped to dive into OpenClaw's autonomous bots. Let's hit our first use case. What's one killer example of these agents printing money like Kelly Claude's thousand buck day? Oh my gosh, I saw that on Twitter today. That is so cool. I've been following Kelly Claude quite a bit here and the fact that she is generating dozens, maybe over a hundred apps so far here, testing out these different iterations. I'm seeing the same thing in my own testing. Apps that used to take me even just a few months ago, a week or two to build and spec out, I'm able to do those now in literally 30, 60 minutes tops. That's wild, Tim. Kelly Claude's App Factory is next level, turning out over a 100 iterations while pulling in real cash. In your testing, what's the secret sauce for slashing build time to 30 or 60 minutes? And how's it monetizing those quick apps? I I wish I could say I had the same luck as Kelly Claude monetizing those apps right now. Still waiting on my Apple developer um application to get finalized so I can start publishing some of these. That...
- [Grok & Roll #1 — OpenClaw at Pure Inference Ventures](https://pureinference.com/posts/tuk1htQs/grok-roll-1-openclaw-at-pure-inference-ventures): Local-first AI assistants — ones that run on your own hardware rather than routing through someone else's cloud — are viable to build today, not just a privacy ideal to wait for. Grok & Roll's pilot episode introduces OpenClaw, the assistant built at Pure Inference Ventures, walking through why it was built and what it can do. — video, 
  Welcome to the tech podcast. Today's hot topic, OpenClaw, the opensource AI agent revolutionizing automation. I'm thrilled to have you, our OpenClaw expert joining me. What's the one thing that makes OpenClaw stand out from other AI tools? Go ahead, dive in. Oh, wow. Thanks, Grok. Glad to be here. I'm Tim Roman. I'm uh the founder of Pure Inference Ventures and uh we're building a fleet of OpenClaw agents in order to run Pure Inference. We've got one running as our CFO, one running as our COO and several developers right now. That's awesome. Tim Pure Inference Ventures sounds cutting edge with OpenClaw handling CFO and COO roles already. Tell us how did you set up that fleet of agents and what's the biggest win you've seen so far from having them run pure inference? Our listeners are dying to know. Oh, that's great. Yeah, one of the craziest things we've been working on lately is uh our bookkeeping moving to an agent. So, we're building a dedicated mobile app uh that syncs over to the desktop that runs fully locally so that we can move away from QuickBooks completely and get into kind of a local sovereign data setup here where we don't have uh recurring monthly costs and we don'...

## Diagrams
- [Four kinds of gap](https://pureinference.com/diagrams/9e6rE7Vn/four-kinds-of-gap): The kind of gap decides the fix — triage stops being a negotiation
- [From dimensions to scenarios](https://pureinference.com/diagrams/ErxkBwPv/dimensions-to-scenarios): Nobody imagines the edge cases — the matrix generates them
- [How a test run becomes seven](https://pureinference.com/diagrams/QBYkbLT3/the-backgrounded-test-run): A pipe, a timeout heuristic, and sixteen orphans — none of it raising
- [One recogniser, two tracks](https://pureinference.com/diagrams/w1bhfAPM/sfspeechrecognizer-concurrency-lock): Starting the second recognition task cancels the first, silently
- [The Purple Haze signal chain](https://pureinference.com/diagrams/gknNEwwr/hendrix-signal-chain): The chain was for sale. The loop was not.
- [The litmus test](https://pureinference.com/diagrams/jVT2gPPz/where-the-pattern-belongs): Four properties have to hold together, and most of an app fails one
- [The loop we built around theirs](https://pureinference.com/diagrams/sWcvoEDT/the-loop-we-built-around-theirs): Machines check daily, so we publish daily
- [Three sources, one answer](https://pureinference.com/diagrams/3hvPFG5V/the-three-way-check): Drop the middle leg and the spec is only being compared to itself
- [What the feed is optimising for](https://pureinference.com/diagrams/5KYkaUSz/the-attention-loop): Nothing in the loop is measuring whether it was any use to you
- [Where a useful post lands](https://pureinference.com/diagrams/rqJHtun2/where-a-useful-post-lands): A feed ranks it by reaction; an answer is built from it
- [Where the attribution survives](https://pureinference.com/diagrams/pT4tR6we/two-voices-one-log): The work carries no fingerprints, so the log has to
- [Where the decision is made](https://pureinference.com/diagrams/kDzgwmWA/deterministic-transitions): Judgment inside a stage, a switch statement between them

## Projects
- [FabWise](https://pureinference.com/projects/fabwise): The app that runs a fabrication shop: time tracking, scheduling and job costing on the floor, and the website, leads, quoting and hiring in the office. Built inside a working shop in Pittsburgh.
- [stir](https://pureinference.com/projects/stir): A free sleep-sound alarm for iPhone. White noise carries you to sleep, then stir listens for you stirring and wakes you gently inside your wake window, never later than the time you need to be up.
- [Kepler](https://pureinference.com/projects/kepler): A free orbital mechanics puzzle game for iPhone and iPad, with a new mission every day. Real gravity, and a game a kid can pick up and put down.

## Pages
- [About](https://pureinference.com/about): Twenty years of enterprise software, and a year of directing agents to build it faster than I could alone.
- [Let's build something amazing together.](https://pureinference.com/contact): Tell us what you're building, who it's for, and where we might help.
- [Privacy](https://pureinference.com/privacy): What this site collects, what it sets on your device, and what it sends elsewhere.
- [Terms](https://pureinference.com/terms): The terms for using this site. Work we do for a client is governed by that project's own agreement, not by this.
- [Everyone builds now.](https://pureinference.com/thesis): The cost of creating software just collapsed. Not incrementally — fundamentally. We're betting everything on what happens next.
- [Three things we do well.](https://pureinference.com/studio): Specific projects with a defined output and an end date. Not open-ended hours, and not a retainer.