Swimlanes — Agentic AI & Operations

02 The work

The rail we actually shipped.

A capture-to-cash billing rail for an SF law firm. Calendar, mail & drive in — QuickBooks out. One automated hop, one human gate.

SL · NOVA-LAW Capture-to-cash

Calendar / Gmail / Drive → portal → QuickBooks

● PRODUCTION 12 wks live

Calendar DWD

Gmail DWD

Drive meta

Manual portal

01 · Ingest

Cron / 30 min

Pulls calendars, dedupes by hash, drops OOO / all-day / reminder noise.

cron · node

02 · Enrich

Draft writer

Turns raw events & threads into billable work-product language.

gpt-4o-mini

03 · Review

Human gate

Attorney edits / approves / rejects. Rate precedence: draft › client › staff.

Next.js portal

04 · Invoice

Rollup

Approved entries across entities roll into one invoice per billing account.

postgres

05 · Sync

QuickBooks

OAuth push of customers & invoices. Tokens encrypted at rest.

qbo · oauth

One autonomous hop. Only the enrich step writes on its own — and only ever into a draft. Everything it produces lands in a review queue.

Nothing bills without a human. No entry becomes an invoice until an attorney approves it. The safety mechanism is the gate, not a promise.

Spine · Postgres — invoicing.*

tenants · staff · staff_rates · clients · client_aliases · billing_entries · draft_entries · invoices · billing_accounts · quickbooks_connections · quickbooks_sync_logs · calendar_event_ingest

Caddy auto-TLS · Keycloak SSO (Google OIDC) + RBAC · Docker Compose · client-owned VPS

Recovered, week one $160k uninvoiced months invoiced after go-live

Daily review ~5min per attorney, replacing a hand-kept workbook

Time to live 4wks kickoff to first invoice through the rail

Access SSO + RBAC Keycloak-backed, fully audited

02.5 Defaults

What we reach for first.

We work in your stack. But left to our own defaults, here’s what we pick — and why. Opinions are signal; “it depends” is not a strategy.

Postgresdata

Multi-tenant schemas, real constraints, boring reliability. We reach for a vector store only when retrieval genuinely needs one — not by reflex.

not Dynamo-first

Next.jsapp + api

One framework for the portal and the API. Server components keep the data next to the render and the surface area small.

one repo

cron → n8n → Temporalorchestration

Start with a cron line. Graduate to n8n when there’s real glue. Reach for Temporal only when durability is the requirement — not the résumé.

simplest that works

OpenAI, Anthropic on failovermodels

gpt-4o-mini handles high-volume text transforms; a codex-class model takes tool calls. Provider order is config, not a wedding. Anthropic is the failover lane.

model-agnostic

Client-owned VPShosting

Your infrastructure account, containerized. We build and operate through the engagement; you hold the keys the whole time.

you own it

03 Practice

Four disciplines. Shipped together.

P / 01

Agent Engineering

Production agents with typed tools, structured memory, and human approval gates where the cost of error is real.

Schema-first

Approval-gated

Your stack

P / 02

Operations Design

Lifecycle maps, SLAs, and explicit handoffs. The unglamorous work that decides whether automation sticks.

Week 1–2

Swimlanes

Documented

P / 03

Evaluation & Safety

Human approval gates where the cost of error is real, transaction-safe validation before anything writes, and a full audit trail on every action.

Approval gates

Rollback-tested

Audit log

P / 04

Team Enablement

Playbooks, not dependencies. Your operators learn to prompt, debug, and evolve the system themselves.

Training

Your IP

Day 90 exit

04 Process

Live in four weeks. Same rail, then maintained.

Short loops. Clear handoffs. Nothing ships without a human gate and a rollback path.

Instrument

Map the workflow. Baseline the metrics.

Week 1

Design

Draft the swimlane. Decide what defers to a human.

Week 1

Build

The system, the tools, the approval gates. A human in the loop from day one.

Weeks 2–3

Prove

Tested, supervised, live. Sign off when it runs clean.

Week 4

Hand over

Playbooks, runbooks, training. You own it.

Week 4

Maintain

Watch for drift. Tune. Evolve. Optional retainer.

Ongoing

★ Field note

The bug we like to tell on ourselves.

In a multi-tenant billing system, every query earns its tenant scope through a join. We never assume a column.

Go-live week, a new draft-count banner counted drafts by a tenant_id column on draft_entries — which doesn’t exist. Tenant scope here flows through the clients and staff foreign keys, never the draft. The page threw a 500 — it failed loud, not as a silent cross-tenant leak.

We caught it in review, rewrote the count to join through clients, and shipped the fix in minutes. No eval harness caught it — a human did. The fix made the rule explicit, and it now rides every query we write.

+ Patterns

The same rail, other back offices.

One engagement is shown in full above. These are representative patterns — the kind of capture-to-action rail we build elsewhere. Numbers are illustrative, not client results.

Illustrative pattern

B2B SaaS · Finance ops

Dunning & collections rail

Overdue invoices in; prioritized outreach and drafted follow-ups out. A human approves anything that reaches a customer; the agent only ever prepares the next move.

~38%faster collection

3 hrs/wk back to ops

1approval gate

Representative scenario · illustrative figures

Illustrative pattern

Field services · Operations

Intake → dispatch rail

Inbound requests in; triaged, enriched, and routed to the right crew out. Edge cases escalate to a dispatcher instead of guessing.

~4×triage throughput

<2 minto first route

0silent auto-dispatch

Representative scenario · illustrative figures

05 Where to start

Not sure where to start? Begin with an audit.

Half-day engagement, $1,500, written report. We sit with the people doing the work, map two or three candidate workflows, and tell you which one is worth automating first — and which one isn't. No commitment to a build.

Workflow map across two or three candidate processes
Ranked recommendation: what to automate first, what to leave alone
Rough scope & price for the recommended build
Written report you own — keep it, shop it, file it away

06 Thinking

AI transformation fails at the operations layer — not the model.

The model works. The system doesn’t. No eval harness. No owner. No runbook.

Swimlanes is built for that gap. Ex-ops leaders, platform engineers, applied-AI researchers.

We leave the build, the tests, and the confidence to change it. Field notes →

Agents that
do real work.