§ — Selected work

Quietly load-bearing software.

A small, anonymized selection. The names matter less than the outcomes — and the outcomes are what we sell.

№ 01

Series-B fintech · KYC / onboarding

Next.jsTypeScriptStripe IdentityPlaidPostgres

Onboarding cut from nine minutes to under sixty seconds.

Problem

A regulated fintech with a 14-screen KYC funnel was losing 38% of applicants between identity capture and bank linking. Their compliance team would not let frontend engineers near the flow.

Approach

We worked alongside compliance to redesign the funnel as a typed state machine, instrumented every step, replaced three vendor SDKs with one, and rebuilt the frontend on a shared design system.

−38%: drop-off in onboarding
9×: faster median completion
0: compliance findings on audit

“Compliance and product, finally on the same screen.”

№ 02

Vertical marketplace · search & ranking

PostgrespgvectorPythonElasticsearchLooker

Search relevance lift that paid for the engagement in a quarter.

Problem

A two-sided marketplace with 1.2M listings had a search experience that hadn't been touched in three years. CTR on results was flat; conversion below benchmark.

Approach

We rebuilt the search pipeline with a hybrid lexical/vector ranker on pgvector, instrumented an offline relevance eval harness, and gave the product team a weekly dashboard.

+41%: click-through on results
+22%: listing-to-booking conversion
0.78: nDCG@10 on the eval set

“Relevance treated like an engineering metric, not a vibe.”

№ 03

SaaS · customer success

ClaudePythonpgvectorLangfuseFastAPI

An AI copilot that paid back its R&D in four months.

Problem

A B2B SaaS support team was drowning in tickets. The CEO wanted “AI somewhere” but the team had seen enough demoware to be skeptical.

Approach

We scoped a single, narrow workflow — drafting first-response replies grounded in product docs and prior tickets. RAG pipeline, eval suite, human-in-the-loop. Everything observable, everything cheap to run.

62%: of first replies drafted by AI
−47%: median time-to-first-response
<$0.04: per draft, fully loaded

“Boring infrastructure dressed up as an AI feature.”

№ 04

Enterprise SaaS · platform refactor

GoPostgresKubernetesTerraformGitHub Actions

A monolith pulled apart without an outage window.

Problem

A 9-year-old Rails monolith was the bottleneck for everything: deploys, hiring, performance. A previous attempt at a rewrite had burned a year and shipped nothing.

Approach

We strangled the monolith service by service against a strict contract. Production traffic shifted incrementally behind feature flags. We did not stop shipping product the entire time.

11 → 2 min: median deploy time
−68%: p95 API latency
0: incidents during cutover

“Modernization without the modernization theatre.”

№ 05

DTC brand · headless commerce

Next.jsShopify Storefront APICloudflareAlgoliaStripe

A storefront that loads before the user notices.

Problem

A growing DTC brand was on a Shopify theme that scored 14 on PageSpeed. Their performance ad spend was wasted on visitors who left before LCP.

Approach

Headless storefront on Next.js + Hydrogen-flavoured commerce primitives, edge-rendered, with a real performance budget enforced in CI.

94: median Lighthouse performance
+27%: paid-traffic conversion rate
1.1s: median LCP, p75 globally

“Speed treated as a product feature, with a CI gate to prove it.”

№ 06

Vertical SaaS · data platform

FivetrandbtClickHouseMetabaseAirflow

From spreadsheets to a single, queryable source of truth.

Problem

A vertical SaaS company had data scattered across Stripe, Hubspot, Postgres, and a graveyard of Google Sheets. Every leadership meeting started with “whose number is right?”

Approach

We built a small, opinionated data platform: Fivetran for extract, dbt for modeling, ClickHouse for analytics, Metabase for self-serve. Every metric defined once, in code, reviewed.

1: definition of MRR
−4 days: monthly board-prep cycle
12+: self-serve dashboards

“Single source of truth, owned by the team that reads it.”

§Begin

Want to be on this page? Tell us.

Start a conversation