Find root cause across any signal your stack emits, from kernel to framework to modem to bus, with findings evidenced against the log lines they came from. Diagnose ships today.
The pain
Below the application layer, system and OS engineering is still done manually.
A CAN race propagating through VHAL into Android Automotive. A modem PLMN drop only on a specific firmware × carrier combination. A Yocto recipe regression after a meta-layer bump that surfaces three boots in. The fix lives in the relationships between signals. Single-layer tools and grep can't see it.
Real device bugs cross kernel, HAL, framework, and modem in a single failure. App-layer crash tools stop at the app. Manual log correlation across subsystems is the only option, and it doesn't scale.
Every device fix runs three to six weeks: triage, root cause, patch, regression, CTS/VTS. Per bug. Dozens per release cycle. Most device engineers spend 40–50% of their time here.
Cross-layer debugging lives in the heads of two or three senior engineers per team. When they're unavailable, issues queue for days. When they leave, the institutional ability to ship goes with them.
What ships today
Upload a bugreport, logcat, dmesg, modem trace, CAN bus capture, or a bundle of them. Get a cited root-cause report back, in minutes.
Multi-step autonomous investigation across bugreports, logcat, dmesg, modem traces, and CAN bus data. Forms hypotheses, correlates events across subsystems, builds cited reports your team can share.
Ask Why did this happen?
Cross-layer multi-file correlation at 1GB+ scale. Compares logs across builds, devices, or time periods to isolate regressions and root causes single-file analysis would miss.
Ask What changed between these two?
Watch Deep Research read one bug report end to end, then return a cited root cause with recommended actions.
Under the hood, four pieces compose every investigation: native parsers across every signal format, cross-subsystem correlation, multi-step autonomous reasoning, and line-level citation back to the source.
Read how Diagnose worksDiagnose works today on a real bugreport. The fastest way to evaluate logcat.ai is to bring us one.
See it on a real bugreportResearch preview
Diagnose finds the root cause. Remediate proposes the patch: cross-layer, cited back to the investigation that produced it, gated by the engineer who has to merge it. AI never commits; the human always does. In research preview now for BYOC (bring-your-own-cloud) deployments.
Outcomes
Engineering-loop outcomes that matter to the people running the release.
Cross-layer investigations that used to span sprints close in a single sitting. The same artifact, the same root cause, an order of magnitude faster.
Senior-engineer judgment encoded in the platform, and the investigations live there, not in two or three people's heads. Any engineer can run what previously required the team's deepest cross-layer expertise, and the institutional ability to ship doesn't walk out the door when those engineers do.
Every finding cites the exact log line, dmesg session, or bugreport section it came from. Failed hypotheses surface as failed, not hidden. Outputs hold up when a chipset vendor, an automotive homologation body, or a customer asks how the conclusion was reached.
Architecture
Citation discipline, hypothesis verification, mandatory human approval, no training on your data. Architectural, not configurable, so the output holds up when a regulator, a vendor, or a partner OEM asks how the conclusion was reached.
Every claim the engine makes cites the source: the exact log line, dmesg session, bugreport section, or device tree node it came from. Findings without citations don't ship. Failed hypotheses surface as failed, not hidden. So reviewers know what was ruled out and why.
The engine forms hypotheses, gathers evidence, and self-corrects when the evidence pushes back. A hypothesis that fails verification is reported as failed. Not hidden, not silently dropped. The audit trail shows what was tried and why each candidate was kept or rejected.
An architectural principle, not a feature flag. Diagnose surfaces cited findings; the engineer decides what to do with them. In research preview, Remediate's AI-proposed patches land under the same rule: the model never commits, the human always holds the merge button.
Customer logs are processed in memory for the duration of an investigation and discarded when it completes. They are never used to train, fine-tune, or improve any AI model, neither logcat.ai's nor any provider's. The zero-training rule is contractual with our frontier-model providers via zero-data-retention agreements, and architectural in our pipeline.
Coverage
Binary formats are decoded natively. Text streams are read without a per-format parser. Recognition is by content, not by file name.
Bugreports · logcat · ANR traces · tombstones · dumpsys
dmesg · journalctl · kernel core dumps · vendor crash dumps · U-Boot
Packet captures · Bluetooth HCI · baseband captures · access point logs
CAN traces · signal databases · automotive traces · robot recordings
System traces · multi-file archives · databases · configuration · test results
Not sure yours is on the list? Upload it. An archive is read as one situation rather than a pile of files, and anything we cannot read is refused at upload with the reason, not accepted and returned empty.
See every supported formatVerticals
Eight verticals where Diagnose ships today.
Cross-ECU fault propagation traced across CAN bus, VHAL, dmesg, and modem in a single investigation.
Vendor text exports (QXDM, Shannon, MediaTek) fused with logcat radio, dmesg, and Android telephony in one investigation. Crosses the chipset proprietary boundary.
Bugreport-shaped Diagnose works day one: system_server crashes, ANRs, HAL failures, framework regressions, OTA-introduced bugs across managed device fleets at scale.
Yocto, Buildroot, OpenWrt, AGL, and custom BSPs across ARM, x86, MIPS, RISC-V. Driver probe-ordering failures, systemd boot issues, recipe-level integration regressions, device tree validation. Works on signals from NXP i.MX, Renesas RZ/G, TI Sitara, Rockchip, Allwinner, and Qualcomm IoT BSPs.
A recording shows the node that stopped publishing; the kernel log beside it shows why. Both are read, on one timeline.
Who manages this device, what policy is enforced on it, and what changed since last month. Read out of what the device already recorded, cited to the line.
An association log names clients by hardware address. Upload the fingerprint table beside it and they become named devices, ranked by how often they dropped.
A failed compliance test names the assertion, not the cause. Upload the result file with the device logs from the same run and they are read together, so the failure points at what to change.
Customer-VPC frontier-model inference means logs never leave your environment. Dedicated cloud instances with custom data retention, SSO/OIDC, and compliance controls. SOC2 Type II certification in progress.
Common questions
The questions device engineering teams ask before pulling logcat.ai into the workflow.
No. Diagnose is autonomous cross-layer investigation: native parsers across every device-software signal format, multi-step reasoning across kernel, framework, modem, and bus, and line-level citation back to the source. Log analysis is the entry point, what an engineer hands the system. The investigation, the correlation, and the cited report are the output.
Diagnose ships today with three investigation modes: Quick (a fast cited answer), Deep Research (multi-step investigation of one capture), and Delta (multi-file comparison) across bugreports, logcat, dmesg, modem traces, CAN captures, and more. Runtime-to-source attribution, our codebase intelligence layer that cites findings back to the platform source that emitted them, is in research preview. Remediate (AI-proposed fixes with mandatory human approval gates) is in research preview for BYOC (bring-your-own-cloud) deployments.
Android: bugreports (zip, 100MB+), logcat in every common format, dmesg, ANR traces, tombstones, dumpsys, event logs. Telecom: QXDM, Shannon, and MediaTek text exports with NAS, RRC, PHY, MAC layer parsing (vendors export binary traces to text via their own tooling; we ingest the text). Automotive: CAN bus traces, VHAL, AAOS logcat. Embedded: dmesg and custom kernel logs across ARM, x86, MIPS, RISC-V. New formats compound the parser corpus. Every customer adds coverage future customers inherit.
Yes. Default deployment is logcat.ai cloud. Enterprise plans include dedicated cloud instances with custom data retention, SSO/OIDC, and compliance controls. Fully self-hosted deployments — BYOC (bring-your-own-cloud) in your own cloud, or on-prem in your datacenter — are available for regulated industries (automotive certification, telecom carriers, defense). SOC2 Type II is in progress.
Custom enterprise contracts with vertical-aware ACVs. An OSS Developer Program exists for open-source maintainers. For commercial deployment, request a technical pilot: we'll bring a real bugreport, run Diagnose on it, and scope the right pilot for your team.
Encrypted at rest (AES-256) and in transit (TLS 1.3). 90-day default retention; configurable on enterprise plans. Log data is never used to train AI models. SSO/OIDC and custom retention policies on enterprise. SOC2 Type II in progress.
Start
Bring us a real bugreport. We'll show you Diagnose on it. From there, we'll scope a pilot for your team.