Fewer claims. Stronger proof.

The group runs some twenty properties on three different property-management systems, an SAP ERP, and more than fifty active tools spread across eight operating layers — the accumulated stack of years of growth, with a provider list to match. Our first deliverables map that ecosystem end to end and define its golden track: which sources should feed the business-intelligence layer first, and which KPIs each system already produces that never reach it — today the BI layer consumes four of those sources, partially.

From there the engagement runs as fractional technology leadership. Automation needs are gathered team by team — finance, operations, expansion — and turned into a prioritized roadmap: the PMS–ERP–CRM integrations that generate the most manual work, the build-versus-buy calls on internal platforms, provider consolidation, and an AI path that runs from guest-facing chat and voice to an analyst agent for evaluating expansion opportunities.

ZenHammer inherited an Expo application built by a previous team between 2019 and 2023. We replaced it. Every organization’s data is scoped by tenant, three client surfaces run against a single GraphQL API, and releases move through a develop → staging → production flow we maintain.

The work spans job scheduling against a resource timeline, Google Maps integration, payment processing with webhook reconciliation, push notifications, deep linking, and over-the-air mobile updates that ship fixes without an app store round trip.

Search runs on three legs fused by reciprocal rank fusion: Postgres full-text for exact matches, trigram for typos and partials, and pgvector for semantic similarity. Organizations and people are classified against twelve controlled vocabularies.

The enrichment pipeline is the distinctive piece. Rather than hard-coding scrapers and prompts, it’s a database-backed, UI-editable library of processes that propose per-field updates for a human curator to review and apply — so the client’s analysts extend the system without engineering involvement. Agentic tool loops run under explicit turn budgets, and every LLM call is traced.

Alongside it: RSS and Atom news ingestion with automatic organization linking, analytics dashboards, static and rule-driven portfolios, geographic mapping with clustering, and full CSV import and export for every entity.

Built and maintained end to end by a single engineer: a monorepo of three services, a claim-and-evidence architecture that keeps each score auditable, and choropleth mapping that renders results geographically.

The product has already been through a generational rewrite — v1 scored individuals, v2 extends the model to places — and carries an investor-facing methodology brief alongside the code.

We were the lead contributor on every repository in the product. Live video runs on AWS Interactive Video Service with IVS Chat for real-time audience messaging, and payments use Mastercard’s encryption and OAuth1 request signing.

The widget is the piece worth calling out: a React application compiled to a standard web component, so a merchant embeds live shopping in their existing theme with a script tag and a custom element — no framework requirements on their side.

LoanWell’s platform is large and long-lived, and its backend belongs to their own engineers. Our remit was the interface — and in lending, the interface is where the difficulty lives: amortization schedules, document generation and editing, dynamic query builders for loan criteria, and spreadsheet-grade data grids.

By the close of the engagement we were also shipping standalone integration widgets connecting the platform to RingCentral and Calendly.

We led the web application — 291 of its 397 commits — and built the browser extension effectively single-handed. The interface pairs two rich-text editing surfaces with a premium data grid, and consumes the client’s own private design system rather than imposing ours.

Repository access runs through a GitHub App rather than OAuth, so permissions are scoped per installation and survive individual users leaving the organization. Each agent run is orchestrated as a container with its own workspace, and logs stream to CloudWatch for inspection after the fact.

The agent layer is deliberately pluggable — the same workflow drives Claude Code or OpenAI Codex through one interface, so the platform isn’t married to a single vendor. Metered usage is billed through Stripe.

Built end to end by a single engineer. The repository history includes commits authored by the platform’s own bot account, which is to say the system shipped some of its own code.

Content is authored once in markdown with a definition extraction and injection system, then ingested by the frontend for learners and by the backend as model context. One source of truth, two consumers — which means the material can’t drift out of sync with what the assistant knows.

Underneath it sits a Terraform-managed AWS environment with Cognito authentication, custom email templates, and a unified-login Lambda handling identity across entry points.

Product Owner, UX Designer, Architect, Developer, Code Reviewer, Security Reviewer, DevOps Engineer, QA Tester, Scrum Master and Release Manager. Each role reads the artifacts the previous phase produced and writes its own, so context moves through the pipeline as documents rather than as one ever-growing conversation.

The part we care most about is that knowledge compounds. Living documents — the PRD, the architecture — carry append-only change logs, while reviews, test results, security reports and sprint summaries land in per-sprint folders. A second run doesn’t start over; it starts from what the first one established. Nothing is pre-scaffolded, so the agents build the project rather than filling in a template.

Written for developers who came to terminal-based agent workflows from VS Code or Cursor and miss the source control sidebar. The file list refreshes live as the agent edits, and selection and scroll position survive the refresh. Ships as a proper herdr plugin with a manifest and integration tests.

Two pull requests, both merged upstream: one bounding the maximum size of keys and values to close a resource-exhaustion path, the other adding the module’s integration test coverage. Small in volume, but merged into third-party infrastructure by maintainers who had no obligation to accept them.

Built for the construction industry as a demonstration of what an always-available bilingual representative could do. The three-column layout puts the editable knowledge base on the left, a live transcript in the middle, and the agent on the right — so a prospect can change a fact and immediately hear the answer change.

A small, deliberately scoped piece of work — included here as an example of turnaround speed rather than technical depth.