Case studies
Proof, not promises.
We are rewriting our case studies around the AI practice. In the meantime, we are happy to walk you through specific engagements on a call, under NDA where needed.
Why this page is thin right now
We would rather show you nothing than show you invented numbers.
Our published work previously covered blockchain infrastructure — wallet systems, exchange platforms and smart contract audits. That work was real and we are proud of it, but it is not what most visitors to this site are trying to buy, and dressing it up as AI experience would be dishonest.
The AI case studies are in progress and most require client sign-off before we can name names or quote figures. If you want detail before then, ask on a call. We can usually describe architecture, constraints and outcomes in enough depth to be useful without breaching anything.
The kinds of problems we are usually brought in on
Four patterns account for most of our work: support deflection, document-heavy retrieval, voice at volume, and internal workflow automation that nobody wants to own.
Support deflection. A team has a growing ticket queue and a knowledge base nobody reads. The naive fix is a chatbot on the docs; the fix that survives contact with real users is a retrieval system with strict grounding, a refusal path when confidence is low, and a clean handoff to a human that carries the conversation context across. The measurable outcome is deflection rate at a fixed satisfaction floor — not deflection rate alone, which anyone can inflate by making escalation hard.
Document-heavy retrieval. Contracts, policies, claims files, technical manuals. The interesting engineering is almost never the model. It is chunking that respects document structure, metadata filtering so a 2019 policy never answers a 2026 question, and an evaluation set built from questions the business actually asks. We build the evaluation set before we build the pipeline.
Voice at volume. Outbound reminders, inbound qualification, collections follow-up. Here the constraints are latency budget, barge-in handling, and what happens when the caller says something the script never anticipated. Indian deployments add language switching mid-sentence, which is a real engineering problem and not a checkbox. See AI voice agent development.
Workflow automation. The process that runs on one person's spreadsheet and three browser tabs. These are unglamorous and they pay back fastest, because the baseline is a human doing something error-prone at 2am. The engineering discipline that matters is idempotency — a retried step must not double-charge, double-send or double-book. See AI workflow automation.
How we agree what success means, before we start
Every engagement opens with a written definition of the number that has to move, who owns that number, and what result would mean we should stop.
Most failed AI projects we are asked to rescue did not fail technically. They failed because nobody wrote down what "working" meant, so the system was judged by demo impressions that shifted with whoever was in the room. We insist on a metric with a baseline measured before we touch anything, and a threshold below which we would tell you to stop spending.
That last part is the uncomfortable one and it is the reason we can be direct about the rest. A vendor who cannot describe the conditions under which they would recommend abandoning the project is not giving you an assessment, they are giving you a quote.
What a reference call gets you
Architecture, constraints, what went wrong, and what it cost to fix — described in enough detail to be useful without naming clients who have not cleared it.
We can usually walk through a comparable engagement end to end: the shape of the data, the retrieval or agent architecture we landed on, the two or three things that broke in production, and the evaluation numbers before and after. Where a client has agreed to speak, we will put you directly in touch with the engineering lead rather than a sponsor — engineers ask each other better questions.
What we will not do is send you a deck of logos we touched once. If a name appears in front of you, it is because that client agreed to it and can be asked about it.
See whether we fit your problem
Tell us what you are trying to build. We will tell you honestly whether we are the right team for it, and what it would realistically take.