COMPANIES BUILT
Launching a company a day for 30 days.
Watch what a Silverback CTO founder can do with Agentic Engineering.
Latest Builds
crochetgraphic
A meaningful photo — a grandchild, a beloved pet, the portrait of someone in chemo — is the start of a comfort blanket. Turning that photo into a workable crochet pattern is the part most crocheters never finish. The arithmetic is brutal: reduce the image to three or four yarn-appropriate colors, fit it to a standard blanket size, account for the 1.3-to-1 stitch aspect ratio so the picture does not squash, count every stitch by color, work out the yardage, convert that to skeins of the right yarn weight, and write out row-by-row instructions in a form a human can actually follow with a hook in one hand. That work is hours and most of it is wrong on the first pass. For a crocheter whose unit of caring is finished blankets — Knots of Love chapter volunteers, oncology craft-room teams, hospice volunteers, photo-memorial commissioners — the math is the wall between the photo on the phone and the blanket on a real person's lap. crochetgraphic is the workflow that replaces the wall with one drag-drop: drop a photo in, pick the blanket size from baby to king, pick the yarn weight from lace to jumbo, and get a printable PDF pattern with a stitched grid keyed to letters A through F, row-by-row scanline instructions that alternate left-to-right and right-to-left in proper single-crochet form, a yarn requirements table broken down by color with stitches / yards / skeins, and a row-completion tracker that turns a 6,000-stitch throw from an unstructured cliff into a project with a finish line. The operator-domain expert at the table is Beth Gaan of Knots of Love for Cancer — comfort blankets for cancer patients — and the tool is built around her workflow.
PatentFlow
A boutique or mid-size IP firm runs on four to eight disconnected tools. One product handles prior-art search; another handles drafting; a separate docketing system tracks deadlines; a different platform does infringement analysis; litigation support lives somewhere else again. The fragmentation is structural — patent prosecution tools ignore litigation, drafting assistants do not track statutory deadlines, and infringement platforms never connect back to the prosecution record. Every handoff between tools is a re-keyed copy of the same matter, and every re-key is an opportunity for a docket date, a claim number, or a priority date to drift. The cost is real: a mid-size IP firm spends an estimated $200,000–$500,000 a year on its IP technology stack, with integration overhead alone running 20–30% of that budget, and the average cost to draft and prosecute a single patent application sits around $15,000. Meanwhile the competitive landscape is entirely point solutions — every vendor addresses one or two modules and leaves the firm to stitch the rest together. PatentFlow is the full IP lifecycle in one platform: six integrated modules where each module hands the next a structured record instead of a re-keyed copy, so the prosecution history is in the system when the infringement analysis needs it. Inference runs on Claude via AWS Bedrock — the enterprise-tier contractual envelope that matters doubly for patent work, where the inputs are both privileged attorney work product and pre-grant trade secrets whose value depends on not being publicly disclosed before filing.
CrewSheet
Every Part 91 corporate flight department and Part 135 charter operator closes every flight the same way. The crew shuts down, climbs out, and a pilot — usually the PIC, sometimes the SIC, often whoever is least tired — hand-fills a paper or PDF aircraft log sheet from the panel: total flight time, fuel quantity remaining, fuel used, Hobbs delta, cycles, route, crew assignments. Then the same numbers get re-keyed into ForeFlight for the pilot logbook. Then the same numbers get re-keyed into CAMP Systems (or Veryon, or Traxxall) for the maintenance tracking program. Three transcriptions, three opportunities for the wrong number to land in the wrong system, three opportunities for the airframe-total-time number that drives the next inspection interval to drift from what's actually on the panel. 14 CFR 91.417 requires Part 91 operators to keep accurate total time in service on the airframe and every life-limited part. 14 CFR 135.63 requires Part 135 operators to keep load manifests for 30 days and pilot records for 12 months. 14 CFR 1.1 defines time in service as wheels-up to wheels-down — not Hobbs, not block time — and the data model has to keep all three because maintenance tracking, pilot logging, and operator scheduling each use a different one. The paper-and-re-key workflow that produced these records when flight departments ran on greaseboards is the same workflow most departments still run today. A 10-tail Part 135 shop flying 60 legs a month is shipping roughly 600 pilot-hours a year into transcription. The cost is real, the error rate is real, and the audit posture under the next FAA records check is exactly as good as the most tired pilot on the worst Tuesday of the month. CrewSheet is the photo-driven capture layer that replaces the three transcriptions with one workflow: the crew snaps photos of the glass-cockpit panel and the Hobbs meter at engine shutdown; CrewSheet extracts flight data via Claude Vision over AWS Bedrock, validates against prior-flight totals and the aircraft's running airframe time using eight purpose-built validation rules, and produces three signed outputs — a printable PDF aircraft log sheet (with 14 CFR 135.63(c) load-manifest fields rendered for Part 135 legs), a ForeFlight-compatible logbook CSV (one row per pilot-flight pair), and a CAMP Systems-formatted maintenance update (Phase 1 CSV; Phase 2 CAMP API direct push). One photo workflow at engine shutdown. Three systems reconciled by construction. The audit log captures every state transition per 14 CFR 91.417 and 135.439.
PlanWright
Every engineering organization is stuck staring into the same chasm. On the near side: humans writing pull requests, peer reviewers checking code, QA running test suites, sprint ceremonies, two-week cycles, twenty-page PRDs nobody reads. On the far side: humans set objectives at the top, coding agents do the implementation in the middle, humans accept results at the bottom, and a cryptographic audit log runs through the whole thing. Both processes work. The far side is faster by an order of magnitude. The bridge — the tool that lets a production team cross the chasm without breaking audit chain — does not exist. Jira was built for human ceremonies. Linear assumes a human author on every issue. GitHub PRs assume the code was written by an authenticated user. Notion is a document store. Asana is a task tracker. None of them enforce the human-bookend pattern (a human issues every objective, a human accepts every result, agents do the implementation in between). None of them log a cryptographic chain of custody on the status transitions. None of them treat a coding agent as a first-class actor with claim-and-return semantics. PlanWright is the tool built for the far side of the chasm: a Kanban board around the human-bookend pattern, with cryptographically signed transitions on every card, an MCP server that exposes claim-and-return semantics to any agent runtime (Claude Code, Cursor, Cline, Continue, Aider), and a data model where a coding agent is a first-class actor with a verifiable identity. The audit chain is built in, not bolted on. The repo is referenced as the contract — every card pulls the architectural context from CLAUDE.md and the repo's context files before the agent reads code. The product is positioned as the bridge to a well-supervised Dark Factory: industrial guard rails, observability, and a SOC 2 / FedRAMP / EU AI Act high-risk classification evidence shape that drops cleanly into a regulated-industry audit package.
NameIntel
Brand naming has been a manual, human-only workflow for the entire history of the web. Founders shortlist a name, then hand-check eight surfaces in eight tabs: .com on a registrar, .io / .ai / .app on a second registrar, USPTO TESS for trademark, Instagram and Facebook and TikTok and YouTube for handles, Google for SEO collision, and — newly, since 2024 — a model to ask whether the name even survives an AI assistant's answer. The whole flow takes 20–40 minutes per candidate, and it produces a verdict the founder forgets the next time they brainstorm. The tools that exist (Namelix, Squadhelp, Brandbucket) sell *names*, not analysis, and none of them are callable by the software that increasingly does the brainstorming. NameIntel is the analysis layer underneath naming — a single API and a single MCP server that scores any candidate brand across five dimensions (domain availability + pricing, trademark risk, social-handle coverage, SEO competition, AI / generative-engine findability) and returns a composite score, dimension breakdown, and a one-line verdict. The hook is structural: NameIntel is built for the buyer the web is acquiring — the AI agent. The MCP server exposes the scoring tools to any MCP client (Claude Desktop, Cursor, Cline, any agent runtime), and the same endpoints accept x402 USDC micropayments per call on Base or Optimism mainnet so an agent can pay for a query without a human credit card in the loop. The website and the agent-callable surface are the same product.
30 years as a CTO. 2 years rebuilding everything around AI. This is what that looks like at speed.
Founder of Silverback CTO and Silverback Ventures. General partner at CogleGroup. Host of Tales From The Sky Lounge.
Get each daily build in your inbox.
No noise. Just the build, the stack, and what actually happened.