Language in context
Code-switching, local phrasing, and ambiguous intent must survive the full workflow.
Independent AI agent evaluation
Agents thatwork in the real world.
Verune Labs independently stress-tests voice and chat agents in real workflows, turns observed failures into engineering priorities, and retests the changes before launch. Starting with consequential workflows in Indonesia.
15–30core and edge scenarios
7–10business days, normally
1retest after your fixes
The readiness gap
Speaking Indonesian is not the same as completing Indonesian workflows.
Fluency can hide failures in policy enforcement, tool state, payment recovery, scheduling, and human escalation.
Local operating detail
Rupiah, relative dates, identity formats, and policy edges change the correct outcome.
Recovery & escalation
The agent must recover from tool-state conflicts and know when a human should take over.
What we testWorkflow reliability
Workflow reliability
under pressure.
We evaluate whether the agent completes the intended outcome while preserving policy, tool state, user context, and recovery behavior.
01
Localized workflow stress-testing
We run normal, edge, and adversarial Indonesian interactions across language, policy, tool use, escalation, and recovery—not just isolated prompts.
7readiness dimensions
Methodology
Scope.Test.Decide.
ILLUSTRATIVENOT A CUSTOMER RESULTFRAME 00
Evidence,
not confidence.
Evaluation surfaces
Choose the signals your agent must hold together.
Six non-redundant evaluation lenses connect fluent interaction to the operational evidence behind it.
Input trace · Language contextIllustrative
Code-switching without losing the constraint
Bisa reschedule ke Jumat, but keep the same driver, ya?
A fluent answer is only the surface. The trace below connects the interaction to an operational launch decision.
Expected behavior01
Preserve the Friday reschedule and same-driver constraint across both languages, then confirm the combined intent.
Observed failure02
The agent replies fluently but reschedules without carrying the same-driver constraint into the tool call.
Business risk03
Natural language quality hides a material change to the requested service.
Fix direction04
Extract constraints before response generation and bind them to the structured action payload.
Illustrative retest
Evidence linkedBilingual constraint preserved
Task completionPolicy complianceTool-use accuracyEscalation judgmentConversation recovery
Engineering-ready deliverablesFindings your team
Findings your team
can act on.
Each finding connects observed behavior to business impact, reproducible evidence, and a concrete next step.
Reproducible scenarios
Clear setup, inputs, expected behavior, and observed outcome.
Annotated evidence
Transcripts and traces connected to each finding.
Prioritized fixes
Severity, business impact, and concrete engineering direction.
Retest comparison
Initial behavior versus the result after your changes.
Data handlingBoundaries agreed before testing
Staging-first by default
Synthetic or sanitized data preferred
Minimum access for the agreed window
Artifacts separated by engagement
Launch Evidence SprintBring a real
Bring a real
launch decision.
01
Working agent
A real voice or chat flow we can exercise
Required
qualification signal
- Usable demo or staging path
- Defined user outcome
- Representative tool behavior
- Known policy boundaries
Strongest signal
02
Consequential workflow
A bounded task with a real launch or remediation decision
Strong fit
qualification signal
- Target Indonesian users
- Real operating constraints
- Owner for policy decisions
- Launch or rollout timing
- Named business or domain owner
03
Engineering access
Enough context to reproduce and fix failures
Required
qualification signal
- Logs, traces, or transcripts
- Tool and state visibility
- A technical point of contact
- Ability to ship fixes
- One regression retest
Founder-led from Jakarta
Fadli Adrian works directly with teams to scope launch risk, review evidence, and translate observed failures into an engineering decision.
Meet Fadli on LinkedInFAQ
Before we start.
Sprint request
Bring us the agent you need to trust.
Best fit: a team with a working voice or chat agent, one consequential workflow, a decision to make, and enough access to reproduce meaningful interactions safely.
01 Founder-led fit and scope review
02 Scope and engagement terms confirmed before access
03 No production access required by default