Skip to content
Selected work

The problems, and what we actually did about them

Most of our work sits behind an NDA, so these are described by shape rather than by logo. The engineering decisions are real, and we are happy to walk through any of them in detail on a call.

A note on these pages. Client identities and commercially sensitive figures are withheld under confidentiality agreements. References are available on request once an NDA is in place.

Digital Asset Infrastructure

Institutional staking platform

The challenge
A digital asset manager needed to open staking to institutional clients without ever holding customer keys in a single place, and without a custody model their auditors would reject.
What we did
We designed a threshold-signing custody layer with policy-based approvals, built the validator orchestration and reward-accounting pipeline, and put the withdrawal contracts through invariant testing and independent audit before mainnet.
  • Solidity
  • Rust
  • Go
  • Kubernetes
  • PostgreSQL
  • MPC

    Threshold custody, no single key

  • 0

    Critical audit findings at launch

  • T+0

    Reward accounting reconciliation

Vertical SaaS

Multi-tenant operations platform

The challenge
A fast-growing operations company had outgrown a single-tenant application: onboarding a customer took weeks of manual configuration and every release risked cross-customer regressions.
What we did
We re-architected the product around tenant isolation with row-level security, shipped a design system and self-serve onboarding, and introduced usage metering wired directly into billing and entitlements.
  • Next.js
  • TypeScript
  • PostgreSQL
  • Stripe
  • AWS
  • Minutes

    Onboarding, down from weeks

  • Self-serve

    Plan upgrades and metering

  • RLS

    Enforced tenant isolation

AI & Analytics

Grounded analytics copilot

The challenge
An analytics vendor wanted a natural-language layer over customer data, but early prototypes hallucinated numbers — unacceptable for a product finance teams rely on.
What we did
We built a retrieval layer over the semantic model rather than raw tables, constrained generation to validated queries, and shipped an evaluation harness that gates every prompt and model change in CI.
  • Claude API
  • Python
  • dbt
  • pgvector
  • Next.js
  • Cited

    Every figure traced to a query

  • CI-gated

    Evals block regressions

  • Sub-second

    Median cached response

Countries served

30+

Countries served

Distributed delivery across US, EU and APAC time zones.

Building since

2020

Building since

Founded as a SaaS and advertising technology company.

Users on shipped products

100K+

Users on shipped products

Across the platforms our teams have launched.

Coverage on retainer

24/7

Coverage on retainer

Follow-the-sun on-call for platforms we operate.

Your project could be the next one

Bring us the problem you have been putting off because it is too intertwined with everything else.

San Francisco, CA · Serving clients in 30+ countries