NorthAnderson

Tools & Integrations

Live components from the systems we build.

Each console below issues a real HTTP request to a public NorthAnderson endpoint running in production. These are reduced, deterministic versions of components we deploy inside client environments — no sign-up, no keys, no simulated output.

Extraction service

Unstructured correspondence into typed, auditable records.

Document classifier

Multi-label taxonomy scoring with evidence and review thresholds.

Schema normalizer

Heterogeneous vendor payloads onto one canonical contract.

Agent run planner

Goals decomposed into ordered, guarded tool-call plans.

Model router

Task profile scored against model tiers with cost and latency.

Routing engine

Deterministic priority, queue, and SLA assignment with signal trace.

Workflow validator

Static analysis of workflow definitions before they ship.

Service status

GET /api/public/v1/status returns the live endpoint inventory.

Extraction & understanding

Reading the documents your systems cannot.

01

Document & correspondence extraction

Operational email, contracts, and vendor documents carry the identifiers your systems need but cannot read. This endpoint returns typed entities — parties, identifiers, monetary values, dates, and contact channels — with per-request coverage metrics suitable for downstream validation.

Use case

A shared accounts-payable inbox receives several thousand vendor messages a month. Extraction lifts purchase order numbers, invoice references, amounts, and dates out of free text so the reconciliation job runs without a human transcribing anything, and every field carries a confidence value the reviewer queue can filter on.

POST /api/public/v1/extract

text
string · required · max 20000
Response
// Response from the live endpoint appears here.
  • Typed entity output
  • Deterministic & auditable
  • Edge runtime execution
02

Document classification

Routing starts with knowing what arrived. The classifier scores a document against an operational taxonomy, returns per-label confidence with the evidence terms that produced it, and flags anything below the review threshold rather than guessing.

Use case

An intake queue mixes invoices, incident reports, contracts, and clinical correspondence. Classification decides which pipeline each document enters, and because the evidence terms are returned with every label, a compliance reviewer can see exactly why a document was routed the way it was.

POST /api/public/v1/classify

text
string · required · max 20000
threshold
number 0-1 · optional
Response
// Response from the live endpoint appears here.
  • Multi-label scoring
  • Evidence terms returned
  • Explicit review threshold

Integration

One canonical contract across every vendor.

03

Integration schema normalization

Every SaaS platform models the same order differently. Paste any vendor payload and the endpoint maps it onto a canonical contract, reporting the strategy used for each field, the coverage ratio, and every source field left unmapped for review.

Use case

A distributor ingests orders from a marketplace API, an ERP export, and two partner EDI feeds. Normalization collapses four shapes into one internal contract, and the unmapped-field report becomes the backlog for the integration team rather than a silent data loss.

POST /api/public/v1/normalize

payload
object · required
source
string · optional
Response
// Response from the live endpoint appears here.
  • Exact, alias & fuzzy mapping
  • Coverage reporting
  • Unmapped field surfacing

Agent runtime

How a run is planned, and what model executes it.

04

Agent run planner

Before an agent acts it should commit to a plan. This endpoint decomposes a stated goal into an ordered sequence of tool calls with explicit dependencies, marks which steps mutate state and therefore need approval, and attaches the guardrails and stop conditions the harness will enforce at runtime.

Use case

An operations team wants an agent to close out vendor invoices. The planner produces the run shape up front — retrieve, extract, reconcile, write, trace — so the team can review and constrain the plan before a single external call is made, and the same plan becomes the replay artifact after the run.

POST /api/public/v1/plan

goal
string · required · max 2000
available_tools
string[] · optional
autonomy
supervised | assisted | autonomous
max_steps
number 2-12 · optional
Response
// Response from the live endpoint appears here.
  • Explicit step dependencies
  • Approval gates on mutations
  • Declared stop conditions
05

Model routing

No single model is correct for every step of a workflow. The router scores a task profile — intent, context size, latency budget, tool requirements, modality — against reasoning, balanced, fast, long-context, and multimodal tiers, then returns the selection, the full score table, a fallback, and the escalation rule.

Use case

A support platform runs three model-backed steps per ticket. Intent tagging goes to a fast tier at sub-second latency, summarization to a balanced tier, and contested escalations to a reasoning tier. Routing this way typically removes most of the model spend without changing what the operator sees.

POST /api/public/v1/route-model

task
string · required · max 2000
input_tokens
number · optional
latency_budget_ms
number · optional
requires_tools
boolean · optional
modality
text | image | audio | mixed
Response
// Response from the live endpoint appears here.
  • Scored tier comparison
  • Cost & latency estimates
  • Fallback and escalation rule

Governance & control

Decisions that must survive an audit.

06

Operational triage & SLA routing

The decision layer inside an agentic workflow: inbound events are scored against weighted signals, then assigned a priority, an owning queue, an SLA clock, and a recommended action set — with the full signal trace returned so every decision can be defended in audit.

Use case

A regulated operator must show why an incident was treated as P1 and why the security officer was paged. The signal trace records each matched pattern and its weight, so the routing decision is reconstructable months later without replaying a model.

POST /api/public/v1/triage

subject
string · optional
body
string · optional
account_tier
standard | growth | enterprise | strategic
systems_affected
number · optional
Response
// Response from the live endpoint appears here.
  • Weighted signal scoring
  • Explainable decision trace
  • Human-review flagging
07

Workflow validator

Most agentic failures are structural, not statistical. The validator performs static analysis on a workflow definition: unreachable steps, transitions to steps that do not exist, mutating actions with no approval gate, missing retry policies, and workflows with no terminal state.

Use case

Before a workflow reaches production it is run through the validator in CI. A payout step with no approval gate and a dangling transition fail the build, which is considerably cheaper than discovering both during an incident review.

POST /api/public/v1/validate

steps
array · required · max 100
steps[].next
string | string[]
steps[].retries
number · optional
steps[].requires_approval
boolean · optional
Response
// Response from the live endpoint appears here.
  • Reachability analysis
  • Approval gap detection
  • Retry coverage scoring

Call it directly

Public endpoints, open to your terminal.

curl -s https://northanderson.com/api/public/v1/status

curl -s https://northanderson.com/api/public/v1/extract \
  -H "Content-Type: application/json" \
  -d '{"text":"PO-48812 issued 2026-08-14 for $184,250.00"}'

curl -s https://northanderson.com/api/public/v1/plan \
  -H "Content-Type: application/json" \
  -d '{"goal":"reconcile invoices and post to the ERP","autonomy":"supervised"}'

curl -s https://northanderson.com/api/public/v1/route-model \
  -H "Content-Type: application/json" \
  -d '{"task":"summarize a support thread","latency_budget_ms":1500}'

curl -s https://northanderson.com/api/public/v1/triage \
  -H "Content-Type: application/json" \
  -d '{"subject":"API outage","account_tier":"enterprise","systems_affected":2}'

Endpoints are rate-limited demonstrations and carry no availability commitment. Production deployments run inside your cloud, behind your identity provider, with logging and controls aligned to your security posture.

Background

How these components fit into an agent architecture.

AI engineering referenceDiscuss an integration