How we engineer

The way we build is the product.

We're a product company that sells engineering, so the practices below aren't a methodology slide — they're how our own software gets made and kept alive. Clients get the same team, the same rules and the same evidence.

How we engineer

Rules we can prove, not values we claim.

These are lifted from the operating manual of our own LLM runtime. Each one is enforced by a test in CI — the code shown is representative of the real thing.

tests/test_guard_01.pyenforced in CI

def test_foreign_token_is_rejected():
    out = observe(page, tokens)
    assert not references_unknown_token(out)  # BLOCKING

Delivery process

From first call to steady state.

Discover

A short, paid discovery: we read your systems and your constraints and write down what 'expensive failure' means for you. You get the document either way.

Contract the work

Scope becomes typed contracts and acceptance criteria before code. Deterministic where it must be, model-assisted where it earns its place.

Build harness-first

The evaluation harness and tests land before the feature. CI is green on every commit; a failing guard blocks the merge.

Ship and operate

Signed releases, rollback by design, observability from day one. We stay on as your operations partner, not a hand-off.

Case studies

Work we can talk about.

Anonymised where the client asked for it. Every outcome listed is one we can substantiate.

Legal & collectionsUS debt-collection law firms

AI redaction that keeps every document inside the firm

FastAPIReactPostgreSQLTesseractpdfplumberWindows edge workerNginx

Custom Software Engineering AI & LLM Engineering

Challenge

Firms file thousands of pages into public court records. Every SSN, account number or date of birth that slips through is a compliance exposure. Manual redaction was slow and inconsistent; cloud redaction tools required uploading client documents to a third party — a non-starter for many firms.

Approach

Built iRedact: a multi-tenant redaction platform with an on-site edge worker so scanned and native PDFs are processed inside the firm's network.
Layered detection: a deterministic net with provable recall on structured identifiers, a model for contextual and OCR-mangled PII, and a foreign-token boundary that rejects any model output not grounded on the page.
Template learning per document layout, jurisdiction-aware rules, and a review lane for anything the two layers disagree on.

Outcome

✓Documents never leave the building; only enum-coded telemetry reaches the cloud
✓Per-page metered pricing with unlimited users
✓Every redaction decision recorded with its evidence and rule version

Legal & collectionsSuitPilot (Goolean product)

Eight applications, one login, one edge node

FastAPINode.jsReactPostgreSQLDockerNginxCloudflarePlaidPlaywright

Custom Software Engineering Cloud & DevOps

Challenge

Collections law firms run a dozen disconnected tools — e-filing, trust accounting, skip tracing, check posting — each with its own login, billing and data-handling risk. Firms wanted cloud convenience without handing over the file room.

Approach

Designed a shared platform: a Portal with SSO, TOTP MFA, Entra ID and per-app entitlements; a unified token wallet for metered billing; and a Goolean edge node on each firm's network that reaches the cloud through signed, outbound-only reverse proxies.
Each application — iRedact, eFile, RemitIQ, Skip Trace, Check Processing, Attorney Review, Analytics — ships as an independent service behind one TLS edge, so a firm can adopt one app at a time.
Browser automation for court portals with no API runs on the edge node and self-updates; bank ingestion via Plaid reconciles trust accounts against firm rules.

Outcome

✓Pay-as-you-go at 1¢ per token, no seats, no subscriptions
✓New applications join the suite without new credentials or new data agreements
✓Firm data stays on-premises; the cloud sees work orders, not files

Legal & collectionsGoolean R&D

Mike: a proprietary LLM that runs inside the firm

PythonApple MLXQwen-VLLoRApytestreportlabTesseract

AI & LLM Engineering Security & Compliance

Challenge

Frontier APIs can't be used on raw client documents, and generic open models can't be trusted to observe a page without inventing what isn't there. Firms needed a model that runs inside their own environment, with proof — not promises — that it never decides, computes or leaks.

Approach

Built Mike as a governed model stack: a domain-tuned vision-language model for document observation, per-firm adapters, and a runtime of typed task contracts, a default-deny policy engine, a hash-chained evidence ledger and an evaluation harness — pre-installed on an appliance delivered to the firm.
Split it into two places: inference on the appliance, and Mike Cloud for signed releases, knowledge packs and fleet health over an outbound-only channel. The cloud never receives a document, a page, a token or a firm's adapter weights.
Every safety property is a test: remote backends refuse raw data, foreign tokens are rejected, deterministic tasks can't be routed to the model, request bodies are never logged. Each guard has a mutation test that proves it isn't decorative.
A synthetic corpus generator produces gold-by-construction documents across scan quality levels and languages; frozen acceptance sets are evaluated once per model version.

Outcome

✓608 automated tests and 43 mutation-tested guards on the runtime (as of September 2026)
✓A 37-task catalog spanning intake, approval, redaction, packet assembly, e-filing and post-filing
✓Runs on a Mac appliance with 128 GB memory; no document or PII ever reaches Mike Cloud

Legal & collectionsUS legal-tech platform

Running a multi-service platform as a managed operation

DockerNginxLet's EncryptCloudflarePostgreSQLGitHub Actions

IT Outsourcing & Managed Services Cloud & DevOps

Challenge

A production suite of customer-facing services — five subdomains, edge workers in customer offices, court portal integrations that change without notice — needed 24×7 operations without a 24×7 US payroll.

Approach

India-based operations team with US delivery leadership; overlapping coverage for US business hours and on-call for the rest.
Runbooks per service, template-skew reviews for the automation layer, rollout checklists with rollback, and monitoring that pages on customer-visible symptoms.
A fixed slice of every sprint spent on reducing toil: automated certificate renewal, self-updating edge clients, audit tooling.

Outcome

✓Single accountable delivery manager; weekly review, monthly SLA report
✓Rollouts with written checklists and rollback on every release
✓Operational cost predictable and materially below an equivalent US team

Tools we ship with

PythonFastAPITypeScriptReactNext.jsNode.jsPostgreSQLDockerKubernetesTerraformNginxAWSAzureGCPApple MLXPyTorchQwen-VLTesseractpdfplumberPyMuPDFTailwindGitHub ActionsCloudflareLet's EncryptAlembicSQLAlchemyVitepytestruffuvPlaywrightPlaidEntra IDPythonFastAPITypeScriptReactNext.jsNode.jsPostgreSQLDockerKubernetesTerraformNginxAWSAzureGCPApple MLXPyTorchQwen-VLTesseractpdfplumberPyMuPDFTailwindGitHub ActionsCloudflareLet's EncryptAlembicSQLAlchemyVitepytestruffuvPlaywrightPlaidEntra ID

Tell us what an expensive failure looks like for you.

One scoping call with a delivery lead and an architect. We'll say plainly whether — and how — we'd take it on.

Talk to us