Outcomes, without the logos.
We don't name our clients, so each engagement below is described by its industry and by the facts we can stand behind. For each one we've also added how we'd run that kind of work with AI today. If you want more, ask on the readiness call and we'll walk you through any of them.
AI operating systems for several clients, in build today.
The work runs from full agentic marketing automation to product development and customer support, on an overhauled technology product portfolio. Ask on the readiness call and we'll tell you what we can.
How AI Business OS works →13× market-cap growth with us as go-to-market lead.
We were its go-to-market lead from 2020 to 2022. In that time we reached hundreds of millions of users, and the company's market cap grew 13-fold.
Read the case →One member record for 70,000 members.
Member data was spread across a legacy portal, an Excel file kept by hand, and email and SMS lists that weren't connected. We mapped all of it, defined a unified record and the rules for it, and handed over the requirements and a roadmap.
Read the case →Four years as one company's technical advisor.
A technical and security retainer that ran for more than four years. It covered security hardening, monitoring and incident response, HIPAA, PCI and SOC 2 readiness, and how the engineering team worked and was organized.
Read the case →A shorter path to a first project.
The client makes a data-labeling platform for machine-learning teams, and its product was confusing new and returning users alike. We ran a heuristic review and redesigned how users create a project.
Read the case →Reading the notes to speed up the referral.
The factors that predict a good specialist referral were already in the physicians' unstructured notes. We scoped the NLP work around that, along with a pilot in three practice areas and an AI roadmap.
Read the scope →More engagements we haven't written up.
These include a speech-recognition company. Ask on the readiness call and we'll go through whichever ones are relevant to you.
We worked with a messaging SaaS company from 2020 to 2022 as its go-to-market lead. Over that window we reached hundreds of millions of users, and the company grew its market capitalization 13-fold. It was acquired for $260M before the engagement ended.
We report these figures as what happened during our engagement. The client stays unnamed, and we're glad to walk you through the work on the readiness call.
This was GTM leadership work, done before today's agent tools existed. We run the same motion with AI now.
Agents handle the groundwork. They research and enrich accounts, score and route leads from your CRM data, draft first-pass outreach for reps to review, and put together the weekly pipeline report. Strategy, pricing decisions and customer conversations stay with people.
Member and program data lived in four places that did not agree: a legacy portal, paper enrollment forms, disconnected email, SMS and event systems, and an Excel file that roughly eight account managers kept by hand as the real working record. Nobody had a shared member record or an attribution model for outreach, and each system named the plans its own way. Every member received the same four emails and four texts a month because nothing could segment them.
We mapped the current state end to end, then designed the target state against it. The unified member record combines eligibility and enrollment, demographics, program participation and engagement scoring, with governance rules that name an owner, an update frequency, a validation step and an audit trail for every field.
We delivered a CRM requirements document, a current-state and target-state data-flow architecture, a data dictionary, and a member-data-flow deck, with a phased roadmap leadership could act on. The roadmap puts a HIPAA-controlled claims and comorbidity analytics phase later on purpose, so that work can slot in without re-architecting the earlier phases.
Automation is only safe once the record is unified, because agents make bad calls on data that contradicts itself.
With the record in place, the next steps are fairly obvious. Segment members by plan and program (instead of sending everyone the same four emails and four texts a month), summarize and tag the free-text contact notes, flag records that break the governance rules so a person can fix them, and give account managers a drafted reply and a member summary before every call.
We keep client names private. On the readiness call we'll tell you what we can about how the engagement was set up.
Four years as one company's technical advisor.
A patient-communications platform in a regulated industry retained us as its technical and security advisor. The work was continuous and covered four areas.
On security, we kept a standing review of the platform, inventoried the API keys, put restrictions on outbound messaging and voice, and formally remediated scan findings. Operations work meant monitoring, post-event log analytics, incident response and the groundwork for geographic redundancy. For compliance we wrote the HIPAA policy documentation and handled PCI and SOC 2 Type II readiness. On the engineering side we set up standard operating procedures, gave documentation and team communication a single home, ran an agile adoption workshop, and split the team into two tiers: a senior architecture group paired with a lower-cost execution group.
Much of this standing work now runs on AI, with a person reviewing it.
AI triages the logs and raises anomaly alerts, takes a first pass at every scan finding, drafts policies and procedures from the controls, and collects SOC 2 evidence as the work happens. That leaves the advisor's time for the judgment calls.
This is how we think an advisory retainer should run. The same advisor stayed in the room for four years.
The client builds a data annotation and project-management platform for machine-learning teams. Its own product was getting in its users' way: the interface gave no clear signal of what had happened, the terminology was dense, the set of options was effectively limitless, returning users got no recognition of having been there before, and tutorial pop-ups kept interrupting people who no longer needed them.
We ran a structured heuristic review of the product and wrote it up against each finding. Then we redesigned the flow that mattered most, creating a new project, and gave the team two shorter options to choose between.
We also recommended instrumenting real usage analytics and running user interviews, so the team could settle the next round of decisions with data on how people actually used the product.
Today we'd put AI inside the same loop.
That would mean a weekly summary of usage analytics and interview transcripts, support questions clustered to show where users get stuck, and an in-product assistant that walks a new user through creating their first project.
The platform connects primary-care physicians to specialists. The working hypothesis was that the time it takes to reach a referral could be cut by making the factors behind a successful referral visible, and that those factors were already sitting in the unstructured notes moving through the system.
Four pieces: cleansing the clinical notes and stripping personal information, a pilot across three practice areas rather than the whole book of business, an exploration of predictive models to score referral notes, and a strategic AI roadmap for what to do with the result.
We're publishing this as a scope, since that's what it was, and there's no shipped result to report. What's worth taking from it is the order of operations: start with the text you already have, prove the signal in three areas, and only then choose a model.
This was AI work from the start, since it meant running language processing on clinical notes.
The same plan moves faster today. Current language models handle de-identification, pull out the referral factors and summarize each note for the specialist, and a clinician reviews before anything changes a referral decision.