Applied Intuition
Each of these is a real decision Applied Intuition is working through, written so you can make the call yourself — incomplete information, someone who pushes back, no answer key.
Applied Intuition is a B2B software company providing platforms to build, test, deploy, and operate intelligent vehicles and autonomous machines. It serves automotive OEMs, commercial fleet operators, industrial equipment makers, and defense organizations, and has strategic relationships with major Japanese automotive OEM customers. Its products are Vehicle OS (including developer tooling and IVI), a flexible suite helping vehicle manufacturers develop, deploy, and update complete vehicle software and AI applications; Tools for Vehicle Intelligence (ADP/APX), for building and validating intelligent systems at petabyte scale; and Self-Driving Systems (SDS). Neural Simulation turns real-world driving data into high-fidelity, photorealistic simulation environments. Its Research Group aims to create technology enabling next-generation physical AI, emphasizing end-to-end autonomous driving and robotic generalists, and can access millions of miles of data from large fleets. It is building a robot learning platform on Dana, its physical AI platform, giving a company the data infrastructure and training intelligence to make any robot learn an industrial task and keep improving it.
Your first one is free. No card.
66 live right now.
The rest
Launch self-service setups for reruns without weakening review gates
Validation engineers are rebuilding cloud setup details before comparable closed-loop campaigns can begin. You have to weigh fast self-service against the controls that make reruns defensible.
Move the 12,000 runs off AWS GPUs without altering evidence
A 12,000-run validation campaign is about to lose its AWS GPU capacity, and paired reruns are not fully matching. You have to weigh a protected review window against a cutover that could alter trusted evidence.
Clear the GPU queue before morning regressions lose their window
GPU simulation jobs are waiting more than twenty minutes when validation engineers need morning regression results. You must weigh faster AWS scale-up against the cost and stability of extra GPU capacity.
Stop rebuilding evidence for known failures while engineers approve
Known Buildkite failures are consuming release-cut time while reviewers rebuild the evidence by hand. You must decide which runbook steps AI can perform and where an engineer must take over.
Package scene edits so release reviews can trust them
Validation reviewers cannot see which parts of a reconstructed scenario came from the source drive. You must weigh broader scenario coverage against evidence a release review can actually follow.
Decide which night evidence to commit before the test window closes
Two physical replays produced near misses after the matching simulated cases passed. You must weigh a fixed approval schedule against evidence that reviewers can defend.
Restore crossing tests before the urban regression review
Pedestrian-crossing replays are landing as uncategorized runs instead of reaching the automated gate. You must weigh a fast metadata fix against preserving every other validation result.
Reconcile Jira and Confluence weekly without creating a second record
Blocked IVI work is being described differently in Jira and the weekly status record. You must decide what the system can reconcile automatically and what remains a program manager’s judgment.
Route OTA alerts before HIL time disappears into repeat runs
OTA validation alerts are forcing repeat HIL runs because key failure evidence arrives scattered across records. You must weigh faster routing against the human review needed before a severity call changes the response.
Trace failed runs while release evidence stays reviewable
Failed closed-loop runs are stalling triage because their evidence is scattered across workflows, logs, and cluster events. You have to weigh faster investigations against the proof engineers need before a release decision.
Keep each run's history when records leave a pilot cell
Robotics engineers are rebuilding policy and evaluation history by hand whenever a pilot run needs review. You have to weigh customer data control against the traceability needed to investigate robot behavior.
Commit to local records for each robot run before cutover closes
Several robot runs cannot be tied to a policy after the factory network drops. You must weigh uninterrupted cell operation against evidence that can support a future policy decision.
Stop robots from running a prior policy after a routine deployment
Approved policies are reaching the edge inconsistently, leaving robots on prior versions. You must weigh fast local deployment against a release path the customer can trust and operate.
Connect mission evidence before the next test plan is written
Program managers reconstruct mission events from separate views and leave reviews without one agreed follow-up decision. You have to weigh a fast shared narrative against preserving the uncertainty in the underlying evidence.
Set reroute authority before the field configuration freezes
Operators cannot tell whether a degraded-link reroute was approved or merely reported. You must weigh mission continuity against an intervention model operators can trust under pressure.
Clarify task state after reconnects without slowing approvals
Operators see an old approval state after reconnecting, even when a vehicle is already executing. You must weigh immediate task clarity against keeping routine approvals fast.
Settle which causes AI can rank while engineers confirm each one
Docker deployments that pass in SIL are failing differently during HIL, leaving engineers to rebuild context from scattered evidence. You must decide what AI can assemble and rank without letting it substitute for engineering confirmation.
Trace failed exports before the next collection cycle
Recorded robot tasks replay correctly, but their training exports cannot be reproduced. You must decide what evidence AI can rank and what engineers must verify before changing the collection workflow.
Link CISPR failures to build records so integration slots survive
Repeated CISPR failures are consuming vehicle integration slots because the test, build, and failure evidence does not meet in one investigation. You must decide what AI may correlate, what an engineer must verify, and how the hardware team runs it afterward.
Prove pilot boards before vehicle slots fill
A 40-unit pilot must reach 95% first-pass yield while each board needs diagnostic evidence before integration. You have to weigh fast board availability against a test method the manufacturing partner can run consistently.
Commit the thermal stack while pilot resets remain unexplained
Compute units restart during sustained thermal runs, putting a 120-unit pilot build at risk. You have to weigh integration timing against committing an assembly path that may carry the failure into every unit.
Fix the harness retention so CAN survives service reassembly
CAN communication fails after service reassembly even when the connector appears latched. You must weigh easier service access against retention that survives repeated handling.
Map SOME/IP onto DDS in one day instead of five
OEM engineers spend two to five days mapping each legacy SOME/IP dependency into a zonal controller. You must weigh reusable defaults against the architecture choices each customer must keep.
Hold or ship the update in 48 hours while traces disagree
Eleven of 4,200 updated vehicles entered a restart loop that left the cluster unavailable. You must weigh a regulatory vehicle milestone against a production defect whose trigger is not yet consistently reproduced.
Fix occupant events without a rollback hitting two other services
Occupant-state events are missing in 18 consecutive virtual ECU starts after a Linux image update. You must weigh a durable SDK correction against a rollback that protects this week’s test slot.
Trace sensor-link failures while flight time stays scarce
Intermittent sensor-link failures are consuming flight-test time because the evidence does not converge quickly. You must weigh faster AI-assisted correlation against the engineer review needed before any change is trusted.
Standardize payload swaps while flight time stays scarce
Every new sensor payload is resetting wiring, Linux configuration, and the route to flight readiness. You have to weigh a reusable interface against OEM fit, serviceability, and the evidence needed before flight.
Choose the aircraft architecture before the flight articles are built
Payload streaming interrupts command-and-control messages during flight, despite clean isolated bench checks. You must weigh a vehicle-side redesign against reducing network demand before the remaining field window closes.
Stabilize navigation before the payload blocks another flight
The vehicle loses PNT/GPS/INS input whenever the payload rail comes on. You have to weigh a fast harness revision against proof that it will survive powered HIL and field checks.
Set a path for variants without splitting the shared baseline
OEM engineers have working regional changes but no common readiness point for integration. You must weigh local delivery commitments against a program structure that keeps the shared platform governable.
Choose a release path for the regional build before branch cut
A regional build passes locally, but the same change removes a startup service from the shared build. You must weigh the OEM integration date against a release split that could persist into later variants.
Set the IVI scope while engineers work from conflicting tickets
An IVI change request is marked for two different releases, leaving engineers unsure what to integrate. You have to weigh customer urgency against the milestone and the work already committed.
Trace release dependencies before OEMs rebuild the evidence
OEM engineers cannot yet see one trusted dependency view across the software they ship. You must weigh broad component coverage against evidence that security can verify.
Set the gate on diagnostic access before the integration window closes
An unauthorized diagnostic request succeeds after some legacy ECU restarts, despite passing clean-boot checks. You must weigh the scheduled integration against security evidence that is costly to reproduce.
Tighten the check on dependency records without blocking legacy updates
An OTA package installed despite one component lacking a complete dependency record. You must weigh stronger release evidence against keeping valid legacy-ECU updates moving.
Rank requests that work offline without widening API access
Offline voice requests have no agreed safe boundary, despite working command demos. You must weigh useful cabin interactions against strict rejection of uncertain vehicle actions.
Commit fatigue alerts while vehicle evidence still disagrees
False fatigue alerts now appear in ordinary daylight driving after the model package changed. You have to weigh the build date against evidence strong enough to sign a safety-bound model.
Recover voice recognition while cabin noise climbs
Noisy-cabin destination requests are missing the recognition target after a model package update. You must weigh a faster accuracy recovery against the latency, memory, and thermal limits of the vehicle hardware.
Surface weak interactions while closed-loop scenes look convincing
Visually convincing reconstructed scenes can respond implausibly when the autonomy stack chooses a new path. You must weigh broad automated coverage against the limited physical evidence needed to trust the score.
Choose highway scenes when physical tests challenge the simulator
Generated highway scenes show fewer collisions, but physical testing does not confirm the gain. You must weigh broader scenario evidence against the correlation needed for a credible validation package.
Fix flicker on vehicle edges in rain without slowing validation
Generated rain scenes raise false vehicle detections 14% above source-log results. You have to weigh visual realism against throughput and evidence the validation engineer can trust.
Build an adoption path while OEM leads still own architecture choices
Individual virtual ECU runs pass, but the OEM cannot connect them to a scale decision. You must weigh a repeatable adoption path against each engineering group’s need to retain architectural control.
Set the service contract before parallel variants become permanent
Two OEM groups have passing virtual ECU runs that support incompatible service contracts. You must weigh platform reuse against the IVI integrations already moving through the customer pipeline.
Rebuild SDK adoption before the next IVI architecture review
Only 38% of enrolled engineers have validated an SDK change through the customer CI/CD pipeline. You must weigh fast onboarding relief against keeping one production-ready validation path.
Expose ARXML gaps without waiting for target integration
Supplier SWCs keep reaching integration with contract gaps that only appear in build output. You must weigh earlier, explainable guidance against the review and configuration boundaries around customer software.
Commit brake recovery while the legacy interface freezes
Brake-state recovery returns a stale value in some restart cycles, and the integration gate closes in five days. You must weigh a compatible interface contract against evidence that is expensive to reproduce on target hardware.
Fix diagnostics status after startup without changing the customer ARXML
The diagnostics status service is missing after target startup, despite the SWC passing in SIL. You must weigh a fast lifecycle configuration change against preserving the customer’s existing RTE contract.
Build weather variants so reviewers stop searching for recorded drives
Pilot reviewers lose days finding comparable fleet drives whenever a regression needs a different condition. You must weigh broad scene control against evidence that each edit remains credible.
Settle where lineage is kept for reruns so outcomes stop conflicting
The same reconstructed drives are producing conflicting pass or fail outcomes for an OEM review. You must weigh replay traceability against capacity, timing, and an irreversible retention change.
Unstick rendered drives before tonight’s validation queue
Only 71% of queued reconstructed logs are ready within 30 minutes, delaying daily regression runs. You must weigh a fast workflow change against preserving consistent sensor outputs.
Surface weak detections without slowing Jetson inference
Troublesome detections are found only after test engineers review broad result sets. You must weigh earlier review signals against latency and false alarms in field preparation.
Choose the IR model for flight test without trading recall for clutter
The revised model improves dim-target recall but increases clutter classifications during HIL runs. You must weigh field detection gains against the tracking and autonomy consequences of a committed release.
Recover detection after handoff while the next field run approaches
Vehicle recall drops to 86% just after sensor-controller handoff, despite strong steady-state results. You have to weigh a fast, reversible deployment change against preserving the accuracy and speed already proven in replay.
Rank interventions so the next policy learns from the right failures
Robotics engineers are reviewing hours of interventions, yet repeat failures still reach the next policy. You have to weigh faster data triage against evidence that the ranking selects genuinely useful failures.
Fix the insertion policy so damaged parts fall below 0.5%
Hard insertions are damaging 3.2% of parts even when the robot marks the assembly complete. You have to weigh a fast return to throughput against evidence strong enough to clear the next policy.
Lift reflective-part picks without slowing the cell
Reflective-part pick success is 82%, and operators are stepping in after visible targets trigger a pause. You have to weigh a targeted policy update against cycle time and intervention risk.
Decide if an interface is ready for HIL when three sources disagree
HIL reservations are being made from separate SIL results, hardware notes, and vendor inputs. You must weigh an early common signal against the evidence needed for engineers to trust it.
Fix timestamp recovery before the last flight window closes
Sensor tracks jump backward after navigation recovery, and the final flight window is already scheduled. You have to weigh usable flight evidence against deploying a change that cannot be casually undone.
Stop old status from reaching operators after a link reconnect
Eighteen percent of HIL reconnect runs show an old task status after the link returns. You have to weigh rapid recovery against the risk of hiding current vehicle state from operators.
Surface bad setups while the robot shift still has time
Missing signals and invalid calibration are being found after recording, wasting scarce robot shifts. You have to weigh an early warning that operators can act on against false alarms that interrupt valid collection.
Decide on action mapping while the next dataset is at risk
Late robot actions could make a 600-episode collection run misleading even when replay looks clean. You must weigh a time-critical cutover against evidence that is expensive to verify.
Fix rejections of good episodes without admitting misaligned ones
Only 76% of pilot-cell recordings are usable because brief timing drift rejects otherwise complete episodes. You must weigh faster recovery of demonstrations against admitting data that cannot support training.
Route AAOS changes through SIL before HIL benches fill
Engineers are consuming HIL slots to find failures that an earlier SIL run could expose. You have to weigh faster self-service feedback against evidence that Program Validation can trust.
Choose the release path while artifact evidence no longer matches
Four signed components differ from their reproducible rebuild dependency records. You must weigh a committed OEM release window against evidence that the binary can be traced and trusted.
Clear the Buildkite queue before integration work slips again
Sixty-three Android Automotive OS pull requests exceeded two hours in the last day. You must weigh faster daytime feedback against capacity reserved for nightly analysis.