The constraint was never the technology
We have spent about twenty years building software and running delivery inside large organizations. The same thing kept happening. The tools meant to help people do careful work made it slower.
Forms that did not match how the work actually happened. Systems that produced an answer and could not tell you where it came from. Reviews that took longer than the thing being reviewed.
The technology was rarely the constraint. The constraint was that nobody had sat with the work long enough to know which part actually hurt.
That is what we do now. We build AI into the tedious, document-heavy middle of careful work, and we build it so someone can check the output.
We are based in Toronto. We work with clients anywhere.
Services
We turn big ideas into reality—minus the corporate fluff, endless meetings, and vague "synergy" talk. We’re here to build, optimize, and streamline, so your business doesn’t just exist—it thrives.
AI application development
We build production applications on Claude. Retrieval over your own documents, structured extraction, multi-step agents, and the evaluation harnesses that tell you when output starts to drift.
That last one gets skipped. If nobody has validated your quality check against real labelled examples, it is not a control. It is a feeling.
AI workflow automation
Automating the repetitive middle of knowledge work. Document review, status synthesis, reporting, the handoffs between systems that people bridge by hand.
We start by scoping what the agent can reach. What it reads, what it can touch, where it can send things. That is a better conversation to have before the build than during your security review.
AI delivery and governance
Four years running delivery inside a multi-year utility program, and about twenty before that. We know how work gets approved, procured and audited where those things are slow and serious.
The usual framing for AI decisions is speed. That is the wrong axis. The axis is reversibility. Safety and data boundaries are one-way doors, so they get built first and in depth. Look and feel is a two-way door, and AI has made it cheap to redo. We sequence the work that way, and it is the reason our projects do not stall at the security review.
What we've built
1 / Droplet Learn | Product | Launched
Turns any goal into a five-minute daily habit. You say what you want to get done. Droplet builds a four-phase curriculum and sends it as daily drops. One lesson, a takeaway, something to try. Built on the Claude API. Free tier and a $9/month plan at dropletlearn.com.
Shipped in 54 days by one person, every commit co-authored with Claude Code. Analytics, billing, error tracking, an admin console and a full design-token rebrand, all before a single public user touched it.
It ships with a five-layer safety system: a cheap pre-classifier on every request before generation, system prompt guardrails, post-generation validation, a public content policy, and safety event logging into an admin console. Generative products can take a user somewhere neither of you intended. There is no rollback on that, so it got built first.
The other hard part was curriculum quality at generation time. Listing subtopics is easy. Producing a sequence that builds, at a sensible pace, for a subject the system has never seen, is the product.
2 / Vellumark | Product | Under Development
AI-assisted claim preparation for SR&ED, the federal tax credit program for Canadian R&D. It reads a company's own engineering record. Commits, tickets, technical notes. Then it drafts the project narratives CRA asks for, with every claim traceable back to what it came from.
Built on what we learned preparing real claims, including where the manual process actually breaks.
Drafting the narrative turned out to be the easy part. Building something whose output survives a CRA review is a different problem. That is the one we are working on.
3 / Agent-assisted delivery workflows | Internal Tooling
Our founder builds and runs agent-based automation for his own delivery work inside a multi-year North American utility program. Document review, status synthesis, and reporting that used to be manual.
This is internal tooling, not a client deliverable. It is also where most of what we know about running agents in a conservative environment came from, including the parts that did not work.