Products we build, ship and run.ship and run.

The brief

Content teams were stitching five tools together and still getting drafts that could belong to anyone. Yahini had to go from a website URL to publishable content without losing the brand on the way.

What we built

• Brand-intelligence agents that crawl a site and turn its product, audience and positioning into structured data every later step reads from.
• A keyword pipeline that scores searches by volume, difficulty and funnel stage, and groups related variants.
• Long-running AI generation jobs with live progress, credit metering and automatic refunds when a run fails.
• A server-rendered React Router 7 app on Cloudflare Workers, with Supabase auth and row-level security on every read.
• A Google Search Console integration over OAuth, and an MCP server that exposes projects to AI assistants.

Under the hood

The brief

Job seekers read a whole posting before finding out it asks for something their CV doesn’t show. Lacuna had to answer “should I apply?” in under a second, without the CV ever leaving their device.

What we built

• A labelling pipeline where an open-weights LLM judges each requirement and Python applies one published rubric, across 12,699 CV and job pairs.
• Synthetic CVs written around real postings at a controlled level of fit, so the model learns strong, partial and poor matches.
• Two encoder models that learn every head in one pass: Altus (395M) as the teacher, and Celer (68M) distilled from it to match it on held-out data.
• ONNX export running on WebGPU through ONNX Runtime Web, checked against PyTorch, with a JavaScript port of the pipeline tested for parity.
• A Chrome side panel that scores the job page you are on and keeps a ranked list of every job you’ve checked.

Under the hood

The brief

On Substack, relationships live in likes, restacks and replies scattered across Notes and posts. Writers had no way to see who actually shows up for them, or who is still waiting on a reply.

What we built

• A Bun and Turborepo monorepo: an Astro web app, a Hono API and a PostgreSQL-backed job runner.
• Ingestion that merges contacts by stable Substack ID and scores inbound and outbound engagement separately.
• Raw API responses kept as evidence, so parser fixes replay without refetching, plus resumable history backfills.
• Row-level security on every tenant table and a restricted runtime database role.
• A Manifest V3 Chrome extension and an OAuth MCP server with scoped grants for AI agents.

Under the hood

The brief

A publication exploring AI, creativity, and culture, alongside a growing library of original tools and experiments. It needed a home that brought everything together without losing the personality that made it distinctive.

What we built

• RobotsOS: tools, skills, agents and playbooks served from one typed catalog, each with an in-place AI chat.
• An OAuth-protected MCP server with its own consent flow, so RobotsOS runs inside the reader’s own AI assistant.
• An admin pipeline that unpacks uploaded ZIP bundles and publishes items straight to R2 storage.
• Supabase auth and protected layouts, rendered server-side on Cloudflare Workers.

Under the hood