Pipeline architecture
Intake normalization (email, SFTP, scanner, API) → layout detection → OCR with vision-language fallback for degraded scans → schema-constrained generation → deterministic validators → reconciliation → write-back.
Services
Seven service lines,
Seven service lines,
assembled around your constraint.
We do not sell a platform and we do not sell seats. Each engagement is scoped from a measured flow, and only the service lines that move the constraint get funded. Everything below is delivered by senior engineers who have run production systems.
SERVICE 01
Flow discovery & process mining
Before anything is built, we establish what the process actually does. We extract event logs from the systems that already timestamp your work — ERP audit tables, ticketing history, mail server logs, MES records, door badges, telematics — and reconstruct the real directed graph of activities, including the paths nobody designed.
The output is not a slide deck. It is a quantified model: activity frequencies, cycle time distributions, queue time between handoffs, rework loop detection, resource contention, and conformance gaps against the process as it was intended.
What you receive
- Discovered process map with volumes and median plus 90th-percentile timings per activity
- Constraint analysis identifying the two or three activities that actually gate throughput
- Flow efficiency ratio — value-added time divided by total lead time, per process variant
- ROI-ranked intervention backlog, including the items we recommend not automating
- Baseline instrumentation so post-deployment gains are verifiable, not asserted
-- Queue time between consecutive activities,
-- reconstructed from the ERP audit trail
WITH ordered AS (
SELECT case_id, activity, occurred_at,
LAG(activity) OVER w AS prev_activity,
LAG(occurred_at) OVER w AS prev_at
FROM event_log
WHERE process = 'order_to_cash'
WINDOW w AS (
PARTITION BY case_id ORDER BY occurred_at)
)
SELECT
prev_activity || ' → ' || activity AS handoff,
count(*) AS transitions,
percentile_cont(0.5) WITHIN GROUP (
ORDER BY occurred_at - prev_at) AS p50_wait,
percentile_cont(0.9) WITHIN GROUP (
ORDER BY occurred_at - prev_at) AS p90_wait
FROM ordered
WHERE prev_activity IS NOT NULL
GROUP BY 1
ORDER BY p50_wait DESC;
-- manual_exception_review is 67% of total lead time
# Durable workflow: survives restarts, retries safely
@workflow.defn
class InvoiceIntake:
@workflow.run
async def run(self, msg: InboundMessage) -> Outcome:
doc = await act.normalize(msg, timeout=30)
fields = await act.extract(doc, schema=InvoiceV3)
# Confidence gate — never guess on money
if fields.min_confidence < 0.94:
return await act.enqueue_review(
doc, fields, reason="low_confidence")
match = await act.three_way_match(fields)
if not match.within_tolerance:
return await act.enqueue_review(
doc, fields, reason=match.variance_code)
# Idempotent write-back, keyed on the document hash
return await act.post_to_erp(
fields,
idempotency_key=doc.sha256,
retry=RetryPolicy(max_attempts=5,
backoff="exponential"))
SERVICE 02
Intelligent process automation
A language model on its own is not an automation. Production workflows need durable state, exactly-once side effects, bounded retries, compensation when a downstream system rejects a write, and a queue that a human can inspect at 2 a.m.
We separate the two concerns deliberately: the model performs interpretation and classification, the orchestration engine guarantees execution. That separation is why our deployments survive an ERP maintenance window instead of silently dropping transactions.
Engineering standards we apply
- Idempotent write-backs keyed on a deterministic hash, so a retry can never double-post
- Confidence gating — any output below threshold is routed to a human, never forced through
- Compensating transactions for multi-system updates that cannot be wrapped in one commit
- Shadow deployment running in parallel with the human process until agreement is proven
- Full replay — every run is reconstructible from its event history for audit or debugging
SERVICE 03
Intelligent document processing
Unstructured paper is where most enterprise throughput is lost. We convert it into typed, validated, reconciled records with an auditable confidence score on every single field.
Accuracy engineering
Per-field confidence calibration against a labelled holdout set from your own archive. We report precision, recall and straight-through rate per document class, and we set thresholds from your cost of error, not from a vendor default.
Human-in-the-loop
Exceptions arrive in a review queue that is pre-filled and source-linked: the reviewer sees the extracted value next to the highlighted region of the original page, corrects it in one action, and the correction feeds the evaluation set.
Document classes we handle routinely
- Supplier invoices, credit notes, statements and remittance advices
- Purchase orders, order acknowledgements, bills of lading, packing slips
- Contracts and amendments — clause extraction, obligation and renewal tracking
- Field service reports, inspection forms, timesheets, delivery proofs
- Insurance claims, medical requisitions, permit applications, grant submissions
- Engineering submittals, shop drawings, technical specifications and standards
Why this outperforms template OCR
Template-based capture breaks when a supplier changes layout, which happens constantly. Modern vision-language extraction reads semantically — it finds the total because it understands what a total is, not because the value sits at fixed coordinates.
We keep the deterministic parts deterministic: arithmetic checks, tax validation, vendor master lookups and tolerance rules are code, not prompts. The model handles perception; verified logic handles correctness.
The result is a pipeline that absorbs new document variants without a configuration project, while still refusing to post anything it cannot verify.
SERVICE 04
Integration with systems of record
Your ERP stays your ERP. We build a thin, well-tested integration layer around the systems you already depend on, because ripping out a system of record to enable an AI project is how AI projects die.
We have worked against modern REST tenants and against twenty-year-old installations with no documented API. Both are tractable — the second simply requires more care around transaction boundaries and rate limits.
Integration patterns in our toolkit
- REST, GraphQL and SOAP clients with contract tests against a sandbox tenant
- Change-data capture on the database journal when no event API exists
- Message-based integration over Kafka, RabbitMQ, SQS or native ERP queues
- Batch and EDI bridges — SFTP drops, fixed-width files, X12 and EDIFACT
- Screen-level automation as a last resort, isolated behind a stable internal interface
- Reconciliation jobs that continuously prove the integration has not drifted
Access model for remote delivery
When a phase runs remotely, these are the accesses we typically request. They are scoped, logged and revoked at engagement close, and we work with whatever your security team can actually approve.
Network
Client VPN or zero-trust broker, IP-restricted
Client VPN or zero-trust broker, IP-restricted
Identity
Named SSO accounts, MFA, least-privilege roles
Named SSO accounts, MFA, least-privilege roles
Data
Read-only replica or API credentials, in-scope tables only
Read-only replica or API credentials, in-scope tables only
Environments
Sandbox or staging tenant for build and testing
Sandbox or staging tenant for build and testing
Telemetry
Audit-trail and event-log exports for the process in scope
Audit-trail and event-log exports for the process in scope
Where remote access cannot be granted — air-gapped plants, classified environments, regulated data residency — the work is performed entirely on-site with your infrastructure.
SERVICE 05
Operational agents & internal copilots
Assistants grounded in your own institutional knowledge, with explicit permissions on anything that changes state. Built for staff who are accountable for the answer.
Retrieval that holds up under scrutiny
Hybrid retrieval combining lexical search with dense vectors, reranking, and metadata filters on effective date, revision, jurisdiction and entitlement. Every answer cites the document and section it came from, so a specialist can verify it in seconds.
We index what your business actually runs on: standards and codes, historical quotes, engineering specifications, SOPs, warranty terms, contract libraries, resolved support tickets and maintenance histories.
Tool use with hard guardrails
Agents that read are low risk. Agents that write require engineering discipline: allow-listed tools, typed parameters, value ceilings, mandatory human confirmation on irreversible actions, and a complete decision log for every invocation.
We define the blast radius before we grant a capability. An agent that can issue a credit note has a dollar limit and a named approver; an agent that can email a customer has a reviewed template set.
Front-line support triage
Classifies inbound tickets, retrieves the resolution history of similar cases, drafts a grounded reply and escalates anything outside its competence envelope with a written rationale.
Estimating & bid support
Reads a tender package, extracts the scope and exclusions, matches line items against your historical unit costs, and flags clauses that deviate from your standard risk position.
Maintenance advisory
Correlates a symptom description with equipment history, manuals and prior work orders, then proposes a diagnostic sequence and the parts likely required.
SERVICE 06
Data foundation & operational telemetry
Most failed AI initiatives are actually failed data initiatives. If the process has no reliable event stream, no model can be trusted and no gain can be proven. We build the minimum viable foundation — not a multi-year warehouse programme.
- Ingestion pipelines with schema contracts and explicit failure handling
- Dimensional and event models built in dbt, version-controlled and tested
- Data quality gates — freshness, uniqueness, referential integrity, distribution drift
- Throughput dashboards your process owners actually open, in Grafana or Power BI
- Model observability — latency, cost per transaction, confidence distribution, override rate
SERVICE 07
Decision intelligence & optimization
Where the constraint is a choice rather than a handoff, the right tool is often mathematical optimization rather than a language model — and knowing which is which is part of what you hire us for.
- Demand forecasting at SKU, location and horizon, with prediction intervals rather than point estimates
- Inventory policy — reorder points and safety stock derived from measured service-level targets
- Scheduling & routing under real constraints: capacity, skills, changeovers, time windows, regulations
- Capacity and staffing models for queue-driven operations such as contact centres and intake desks
- Pricing and margin analytics with explicit sensitivity to cost drivers
CROSS-CUTTING
Governance, security & compliance
Applied to every engagement, not sold as an add-on. An automated decision you cannot explain to an auditor is a liability, however accurate it is.
Traceability
Every automated decision stores its inputs, the model and prompt version, the confidence scores and the rule path taken. Reconstructible months later, on demand.
Data minimization
Personal information is redacted or tokenized before inference wherever the task does not require it. Retention windows are configured per data class.
Residency & hosting
Cloud regions in your own jurisdiction, your own tenant, or fully self-hosted inference on your hardware with open-weight models when data cannot leave the building.
Regulatory alignment
GDPR and the EU AI Act, PIPEDA and provincial privacy law in Canada, US state privacy regimes, plus sector requirements such as SOC 2 evidence support and ISO 9001 process documentation.
Sectors
If there is IT in the business, there is flow to recover.
Our method is sector-agnostic because constraints are structural. What changes is the vocabulary, the systems and the regulatory envelope — not the physics of a queue.
Manufacturing & industrial
Scheduling under changeover cost, quality inspection triage, MES-to-ERP reconciliation, predictive maintenance intake, supplier quality correspondence.
Distribution & logistics
Demand forecasting, replenishment policy, carrier document processing, exception handling on shipments, dock scheduling, claims recovery.
Professional services
Accounts payable and receivable automation, engagement intake, document assembly, time capture from activity data, compliance checklists.
Construction & engineering
Tender analysis, submittal review, RFI triage, change-order tracking, site report digitization, progress claim assembly.
Healthcare administration
Referral and requisition intake, prior authorization packaging, billing code validation, denial management, scheduling optimization.
Finance & insurance
Claims first-notice triage, underwriting document review, KYC evidence assembly, reconciliation breaks, regulatory reporting preparation.
Retail & e-commerce
Tier-one support deflection, catalogue enrichment, returns disposition, marketplace dispute handling, fraud review queues.
Transportation & field service
Dispatch optimization, proof-of-delivery capture, driver document compliance, work order enrichment, warranty claim validation.
Public sector & non-profit
Application and permit intake, grant reporting, case file summarization, records requests, eligibility pre-screening with full auditability.
Commercial questions
Straight answers.
On-site is our default, particularly for discovery — you cannot measure a flow you have not watched. Engineering and industrialization phases are frequently delivered remotely once we hold the necessary system access, with on-site checkpoints at each rollout gate. Where remote access cannot be granted, the entire mandate runs on-site. We define the split, the required access rights and the revocation date in the statement of work.
For discovery: read-only access to the event and audit tables of the systems carrying the process, plus exports of relevant mailboxes or queues. For build: a sandbox or staging environment, scoped API credentials, and a named SSO account with least-privilege roles. We never ask for shared administrator credentials, we operate through your VPN or zero-trust broker, and all access is logged on your side and revoked at close.
That is a decision for you, not an outcome we engineer toward. What we consistently remove is transcription, chasing, re-keying and exception triage — work that nobody was hired to enjoy. In most engagements headcount is redeployed onto analysis, customer relationships and the exception cases that genuinely require judgment. We say this plainly because adoption fails when the people operating the process believe the project is aimed at them.
No. Every operational dataset we have encountered was messy, and the discovery phase exists partly to quantify how messy. If the data foundation is genuinely insufficient to support the intervention, we will tell you that in the diagnostic report and scope the foundation work separately rather than building on sand.
You do. Deliverables include source code in your repository, infrastructure definitions, runbooks and evaluation datasets. We deliberately avoid proprietary lock-in: you must be able to replace us without losing the solution, and we consider that a design constraint rather than a concession.
The flow audit is a fixed-fee, fixed-scope engagement with a defined deliverable set. Build phases are quoted per milestone against the acceptance criteria agreed at the end of discovery. Where the baseline is measurable and the client prefers it, we will structure part of the fee against verified throughput improvement.
Then the report says so. A meaningful share of the constraints we find are resolved by a database index, a corrected approval threshold, a removed redundant sign-off or a fixed integration — and we will recommend the cheap fix over the interesting one. Our credibility depends on it.
Which flow is costing you the most?
Tell us where the work piles up. We will tell you whether it is measurable, addressable, and worth the investment.