RETEST·ASIA — AI ENGINEERING STUDIO

We build the systems
your competitors
call magic.

AI agents, LLM harnesses and automation pipelines engineered around your data, tools and operating rules, with boundaries agreed before implementation.

Status: taking project enquiriesLAYOUT: MEASURING…

Scroll to enter the network

AGENTS

Bounded actions

RAG

Cited retrieval

FLOWS

Queues + retries

APPS

Operational software

THESIS.01 — SIGNAL

/synthetic/intake — SYNTHETIC TRACE

request received · language detected context retrieved · source access checked route selected · confidence recorded uncertain cases stop for review

/synthetic/contracts — SYNTHETIC TRACE

relevant clause retrieved · source attached related clause retrieved · scope compared low-relevance passage removed signature material skipped

/synthetic/evals — SYNTHETIC TRACE

faithfulness check · passed retrieval coverage · reviewed regression against baseline · clear

retriever.topk = configured rerank.model = selected · shortlist kept cache.ttl = policy · batch = bounded escalate.threshold = evaluated

/synthetic/operations — SYNTHETIC TRACE

manual handoff identified in the workflow agreed: system drafts, operator approves every action keeps a trace · kill switch stays

q: what did we agree about this delivery? grounded: approved sources · confidence recorded cited: agreement, correspondence, call notes

/synthetic/schema — SYNTHETIC TRACE

table shipments(reference, eta, carrier, status) table invoices(reference, vendor, total, due_at) index invoices_due_at · configured

alert: extraction drift on vendor invoices template rollback completed · resolved postmortem filed · eval case added

/synthetic/crm — SYNTHETIC TRACE

account record · renewal note present account record · risk flag reviewed account record · follow-up note queued

the retriever returns twelve candidates; the reranker keeps four; one is cited. most of a corpus is noise on purpose.

/synthetic/pipeline — SYNTHETIC TRACE

retrieve · approved documents returned draft · sources attached · grounded evaluate · pass · action: approved

input budget · within configured limit cost per run · recorded fallback route · available

/synthetic/discovery — SYNTHETIC TRACE

the bottleneck is not the model. it is the four systems that never talk, and the person retyping between them.

chunking: by clause, not by page. citations required · answers refuse without a source · uncertainty routes to a person.

/synthetic/inbox — SYNTHETIC TRACE

supplier eta changed · order reference attached draft reply prepared · awaiting approval crm update held until approval

eval case: ambiguous refund clause expected: escalate · got: escalate · pass threshold behaves as configured

/synthetic/access — SYNTHETIC TRACE

scope: read crm, draft email, no send approval: required for sensitive action audit: calls logged · retention configured

ingest approved files · permissions retained dedupe complete · ocr routed where needed answer returned · sources attached

The model can read the corpus. The work is pointing it at the right thing.

The model is not the strategy. We identify the decision, evidence and action that matter — then engineer the system around them.

SYS.02 — Loop engineering

A model call is not
a production system.

Chatbots fail when the model is allowed to improvise past the evidence. We build measured loops that retrieve context, constrain tools, test each output, correct failures and escalate uncertain cases to a person.

01RETRIEVE02RUN03EVAL04CORRECT

PASS 1 — DRAFTEVAL: FAIL

RETRIEVE · plans corpus → 3 chunks

The enterprise plan includes SAML SSO and SCIM provisioning. Setup is instant and enabled by default for every region.

FAIL — claim 2 unsupported by retrieved context

PASS 2 — CORRECTEDEVAL: PASS

RETRIEVE · cited → plans.md §4.2

“Enterprise SSO is enabled after a security review — typically two business days.”

The enterprise plan includes SAML SSO and SCIM provisioning. Setup requires a security review before SSO is enabled — typically two business days [plans.md §4.2].

PASS — 2/2 claims grounded in cited context

FAILURE ROUTE: HUMANCONTROL: KILL SWITCH

SYS.01 — Capabilities

See the system, not the slogan.

Four representative architectures show how we turn an operating rule into a bounded, observable production system.

SYSTEM PATTERN / SELECT A BUILD

One constraint. One observable path.

These are representative patterns, not client results. Each build is designed around the real data, permissions and failure modes in scope.

REPRESENTATIVE ARCHITECTURE

NO CLIENT DATA

01 / INPUT

A qualified prospect replies with a product and pricing question.

02 / DECISION PATH

01Classify intent
02Read account context
03Check response policy

03 / TOOL + ACTION

Draft the reply and prepare the CRM update.

05 / OUTCOME

An approved reply is queued and the opportunity trace is stored.

FULL CAPABILITY LEDGER

Seven ways we build.

The architecture changes with the constraint. The engineering standard does not.

01
AI agents
Agents that read, decide and take bounded action across your inbox, CRM and internal tools — with approval gates and a trace for every run.
tool use / approval gates / traces
02
LLM harnesses
Control layers around models: routing, structured outputs, fallbacks, caching, observability and evals.
evals / routing / fallbacks
03
Automation pipelines
Deterministic workflows with queues, retries and alerts that move data and work between systems without manual handoffs.
webhooks / queues / retries
04
Second brains
An org-knowledge product that makes documents, decisions and internal expertise searchable through one permission-aware interface.
knowledge graph / search / access control
05
Production-grade RAG
Retrieval infrastructure that grounds answers in your source material, with deliberate chunking, citations and evals for recall and faithfulness.
grounding / chunking / citations
06
CRM/ERP web apps
Operational software shaped around your workflows, roles and reporting — not a template your team has to work around.
workflows / permissions / dashboards
07
Conversion pages
Fast, instrumented pages that explain the offer, remove friction and turn qualified attention into a measurable action.
webgl / gsap / analytics

PIPELINE / 06 STAGES

SYS.02 — Process

From brief to a production plan.

We study how your business actually runs — people, systems, data and edge cases — before choosing an architecture. Then we combine that operating context with current AI and engineering practice to build the correct system, not a generic one.

01Study the business

We interview operators, observe the workflow and inspect the systems, data and exceptions that shape it.

/DISCOVERY · WORKFLOW MAPPED

02Map the right system

We define the decisions, inputs, actions and boundaries — then determine where AI belongs and where deterministic code wins.

/SYSTEM-MAP · BOUNDARIES SET

03Set the evals

We turn real cases and failure costs into test sets, acceptance thresholds and escalation rules before the build begins.

/EVAL-DESIGN · CASES SET

04Build and connect

We implement the agents, retrieval, pipelines and interfaces against your stack in a workspace you can inspect.

/BUILD · CONTROL LAYER CONNECTED

05Break and harden

We red-team edge cases, test permissions and fallbacks, and close reliability and latency gaps before launch.

/HARDENING · FAILURE ROUTES TESTED

06Ship and watch

We deploy with traces, alerts and a kill switch, then monitor the launch support window agreed for the build.

/PRODUCTION · RELEASE REVIEWED/SUPPORT · WINDOW AGREED

01Study the business

We interview operators, observe the workflow and inspect the systems, data and exceptions that shape it.

/DISCOVERY · WORKFLOW MAPPED

02Map the right system

We define the decisions, inputs, actions and boundaries — then determine where AI belongs and where deterministic code wins.

/SYSTEM-MAP · BOUNDARIES SET

03Set the evals

We turn real cases and failure costs into test sets, acceptance thresholds and escalation rules before the build begins.

/EVAL-DESIGN · CASES SET

04Build and connect

We implement the agents, retrieval, pipelines and interfaces against your stack in a workspace you can inspect.

/BUILD · CONTROL LAYER CONNECTED

05Break and harden

We red-team edge cases, test permissions and fallbacks, and close reliability and latency gaps before launch.

/HARDENING · FAILURE ROUTES TESTED

06Ship and watch

We deploy with traces, alerts and a kill switch, then monitor the launch support window agreed for the build.

/PRODUCTION · RELEASE REVIEWED/SUPPORT · WINDOW AGREED

ARCH.01 — THE EDGE

Models change. Your system should not need a rebuild.

We engineer the layer around the model so your system can adopt better models without losing its data, logic or operating history.

A swappable model layer

Prompts, tools, retrieval and evals stay outside any one provider. New models can be benchmarked and adopted without rewriting the workflow.

PROVIDER: SWAPPABLE

Upgrades earn their way in

We test each candidate against your eval set for quality, latency and cost before production traffic moves.

GATE: EVALS PASS

Ownership is defined up front

We document the ownership, deployment boundary and provider dependencies for each build before implementation begins.

OWNERSHIP: AGREED

SEC.01 — SECURITY

Security is the first constraint.

Agents touch real systems and can take real action. We design the boundary before we grant access.

Your deployment boundary

We agree the deployment boundary for each build before access is granted. Data stores and credentials stay within the approved architecture.

BOUNDARY: AGREEDDATA: SCOPED

Least privilege, reversible action

Every agent receives only the tools, records and actions required for its job. Sensitive actions require approval; kill switches can revoke execution immediately.

ACCESS: LEAST PRIVILEGECONTROL: KILL SWITCH

Auditable by default

Tool calls, retrieved sources, outputs and approvals can be logged with configured redaction and retention controls. Provider training and retention settings depend on the selected providers and contract.

TRACES: CONFIGURABLERETENTION: PROVIDER-DEPENDENT

PATTERN.01 — System patterns