Enterprise Software Development, with the architecture decisions shown

The challenge

A checkout replatform had been in progress for eleven months with no production traffic. The in-house team was carrying the legacy system’s incident load and could not get a clear run at the new one. Conversion on mobile was eleven percent below the sector benchmark and the cause was unknown.

What we built

• Dedicated team took full ownership of the checkout domain including its on-call rotation
• Progressive rollout by traffic percentage and market, with automated rollback on conversion regression
• Payment orchestration layer abstracting three regional providers behind one idempotent interface
• Real user monitoring wired to the conversion funnel, making the mobile regression visible within a week
• Contract tests between checkout, pricing, inventory and fulfilment, removing the shared staging bottleneck

The challenge

Clinicians were searching four disconnected repositories of protocols, formulary rules and local policy, averaging six minutes per lookup. An earlier vendor pilot had been withdrawn after clinicians found confidently worded answers that cited protocols which had been superseded years earlier.

What we built

• Document ingestion with version-aware chunking so superseded guidance is never retrieved as current
• Hybrid retrieval - dense vectors plus BM25 - with a cross-encoder reranker tuned on clinician-labelled pairs
• Golden set of 640 real clinician questions, annotated by practising staff, executed on every pipeline change in CI
• Mandatory citations with deep links to the source paragraph, and an explicit abstention path
• Entirely inside the customer VPC, with private model endpoints and no data leaving the compliance boundary

The challenge

The carrier portal ran on a single shared Postgres schema with tenancy enforced in application code. Two cross-tenant data exposures in eighteen months had reached the board. Peak-season load produced 40-second page loads, and the largest twelve customers had contractual data residency requirements the platform could not meet.

What we built

• Tenancy moved behind a single data-access seam, then to row-level security with transaction-scoped tenant context
• Dedicated database tenancy for the twelve largest accounts, behind the same routing layer
• Next.js front end with server components and streaming, replacing a client-rendered single-page application
• Read models projected from the operational store for the reporting views that caused the worst load
• Load testing wired into the release pipeline with peak-season traffic profiles

The challenge

Loan origination ran on a COBOL core with overnight batch settlement. New products took nine months to launch because every change required a coordinated release across four teams, and the two engineers who understood the batch scheduler were both within three years of retirement. Regulatory reporting was assembled by hand from four sources each quarter.

What we built

• Routing seam in front of the core, with per-product traffic rules flippable without a deploy
• Change data capture from DB2 into a Kafka log, giving new services a replayable event stream
• Product-by-product extraction into Java services with Postgres ownership, dual-written and reconciled hourly
• Event-sourced ledger for settlement, replacing overnight batch with continuous posting
• Automated regulatory reporting derived from the event log rather than from nightly extracts