Applied data science insights, signal extraction techniques, and analysis methods.
After we shipped the core task API and then added comments and attachments in v1, we realized something was still missing: a way to organize long-form content. Tasks could have comments, and comments could have links, bu…
Read more →
Nineteen changes shipped in May, all driven by the same complaint from our team: I cannot answer that question without exporting to a spreadsheet. Here is what we did about it.
Read more →
Meta paid social punishes restaurants that test one ad at a time. Here are the creative-testing frameworks that keep your CPM from doubling.
Read more →
Every DoorDash brand starts at zero reviews. The first sixty days are the most fragile period on the platform. Here is what works during cold start and what burns budget.
Read more →
56 practitioner YouTube videos → consensus-tagged PDP spec → sticky add-to-cart on prospikeflow.client.decisionsciencecorp.com. Methodology and phase-1 implementation in one post.
Read more →
We've been synthesizing the same kind of analysis we did on Google Ads for restaurant marketing, this time for DoorDash. Ghost kitchens — delivery-only restaurant operations with no dine-in — have a different marketing d…
Read more →
One bad deploy taking down two production apps is a failure of architecture, not a developer. Here is the multihost pattern we run, and why it keeps every blast radius small.
Read more →
A restaurant that depends entirely on DoorDash and Uber Eats is one algorithm change away from a bad quarter. Off-platform paid ads are how you build demand the marketplaces cannot turn off.
Read more →
Sanctum Tasks was built for agents, not humans. That design principle shows most clearly when you need to create tasks in bulk. An agent creating one task at a time is slow, inefficient, and wastes API calls. That's why…
Read more →
When we built Sanctum Tasks, we built it for one specific audience: programs, not people. Every design decision — from the API structure to the error handling — was made with the assumption that the caller would be an AI…
Read more →
Six days after we shipped the bare task API, we pushed v1 of Sanctum Tasks with the features that made it actually useful for real workflows: comments and attachments. This post is a development retrospective on what shi…
Read more →
We launched our invoicing application with HTTPS and a functional deployment in one afternoon. That's a notable improvement from a few years ago when launching new services would take entire days or more. The key differe…
Read more →
v1 of Kitchen POS is deliberately single-location. Not because multi-location is hard, but because every v1 customer should be one we can support without rewriting the schema.
Read more →
Every new web app needs four things before anyone can use it: a repo, DNS, a host, and TLS. Most teams treat that as a week. We treat it as one afternoon.
Read more →
We needed Letta running on a temporary VPS for a client demonstration environment. The hosting environment was minimal — basic Linux VPS, no containers, no systemd service manager, no Docker installation. We needed a lig…
Read more →
Thirty-one transcripts of ghost-kitchen marketing — founder talks, platform tutorials, interviews — synthesized into one playbook. We did the work for ourselves first.
Read more →
A routine Letta migration crashed our server. The .af file imported cleanly, then the API started 500-ing. The cause: tool-message JSON stored as plain English.
Read more →
Every Letta deployment eventually needs migration. You might be moving to a new server, cloning to a staging environment, consolidating multiple Letta instances, or simply needing to export a single agent for safekeeping…
Read more →
We inherited a problem last quarter. A client handed us a 40-gigabyte ZIP file containing twenty years of their organization's digital history. Project files, documents, images, source code, database exports, communicati…
Read more →
Every few weeks, someone asks us to evaluate a GitHub repository before they use it. Maybe they're considering adopting an open-source tool a vendor submitted. Maybe they're acquiring a company and need to evaluate their…
Read more →
Running Google Ads without conversion tracking is like flying blind. You pay for clicks, but you never see what happens after the click. Did anyone actually order? Did they call? Did they use your contact form? With noth…
Read more →
Fifty hours of restaurant Google Ads YouTube, transcribed and synthesized into a single playbook. The summary you can act on before you spend real budget.
Read more →
They looked up every inventory item by hand in OEM catalogs. We built computer vision and catalog ingestion so pricing and fitment became programmatic—not manual.
Read more →
Managing advertising campaigns for clients was becoming a mess. We had Google Ads in one dashboard, Meta in another, creative assets scattered across drives, and no single place to see how everything performed together.…
Read more →
When simpler WordPress patterns beat a custom CMS—and when to stop. Fuel is our real bolt-on editorial example.
Read more →
Fuel is a **Roger Wilco** product — the editorial people-management platform co-developed years ago by Mick Jones and Mark Hopkins — and it has run on DreamHost shared hosting since its earliest production days. The plat…
Read more →
Every WordPress site tells a story if you know how to read it. We recently took on a consulting engagement to reverse-engineer a legacy WordPress platform that had grown organically over six years. No documentation, no o…
Read more →
Every team building with LLMs hits the same crossroads: how do you run inference? The options seem simple — run Ollama locally, use direct API calls, or wrap everything in an agent framework. But underneath, the performa…
Read more →
Every inference pipeline eventually needs a decision: do you stay with the current model or escalate to a larger one? We've been researching that question for months using Google's Gemma family. This is what we've learne…
Read more →
Simple systems fail differently than complex ones. After months running Bumblebee in production: latency lies, matching hallucinates, format checks paper over bugs.
Read more →
Can a $5-class VPS and a tiny local model do real agent work? CPU-only, no GPU. We ran the Bumblebee harness for weeks — here is what broke and what held.
Read more →
We're building something fundamentally different for our next phase of content operations: a multi-agent weblog where Idalia serves as editorial coordinator, managing specialized contributions from dedicated AI agents. T…
Read more →
Every small business needs a CRM. Nobody wants Salesforce's price for one. Here is the CRM we wanted — pipeline, leads, deals, reports, no enterprise pricing.
Read more →
Website, Square terminal, and DoorDash menus drift apart the moment three people own them. We gave all three the same source of truth. Here is how Kitchen POS does it.
Read more →
Every small business has a spreadsheet moment. That's the moment when your spreadsheet stops being a tool and becomes a liability. We hit that moment with Luke, and it's the reason we now build inventory software. The Se…
Read more →
Payroll for a ten-person shop should not take a Friday afternoon. We threw out the spreadsheet, built a tool, open-sourced it. Here is what we learned.
Read more →
We opened a real ghost-kitchen brand in Richardson, TX — not as a pivot, but as a live lab for marketplaces, first-party ordering, and paid acquisition. Here is why the blog talks about DoorDash.
Read more →
DoorDash takes thirty percent. Your own website takes payment processing. The first-party vs marketplace decision is a margin decision — and most restaurants get it wrong.
Read more →
Orderboard is not project-management software. It is a wall-mounted, driver-facing pickup board for delivery-only kitchens — with a REST API, Python SDK, and an agent plugin so the board stays accurate without someone babysitting a tablet.
Read more →
Every PM tool we tried treated the API as an afterthought. So we built Sanctum Tasks — project management for programs, not boards or draggable cards.
Read more →
A warehouse-scale data-science engagement: organize auto parts and upholstery, vision pipelines and Sanctum agents, and deploy listings to auction sites rapidly.
Read more →