Every deploy
changes the logs.
We're building small, log-native models that learn with your software. Inside your cloud. From every line and every on-call decision.
Logs change.
The model should, too.
-
01
Read every line.
Small models, built for logs. Designed to run on CPUs inside your cloud, close to the data.
-
02
Learn the new normal.
Adapt to new formats and behavior from unlabeled logs as teams deploy new code.
-
03
Learn from on-call.
An engineer's confirm or dismiss becomes a signal for the next model update.
The difference is where you start.
A release can introduce a new failure or simply log an old event differently. Start with a comparison, then inspect the original lines behind every pattern.
Explore a release investigationrequest_id=req-98 duration=320ms status=503
Keep the engineer in the loop.
A useful signal needs evidence. Trace patterns to source lines, record what you checked, and keep or dismiss the finding. The workspace preserves that decision with the investigation.
Review an investigationContinuous learning needs a safety net.
Test before an update ships.
Our design calls for every update to be replayed against past incidents. Learning a new normal must not erase the ability to recognize an old failure.
Make every version reversible.
Versioned updates and rollback are part of the model lifecycle we're building. Changes should be measurable, reviewable, and reversible.
Keep a short evidence trail.
The goal: catch novel behavior before it becomes a page, with cited log evidence that engineers and AI agents can inspect.
A team close to the problem.
Karan Bhukar
Researching distribution shift and out-of-distribution coverage at NYU. His BERTOps work learned from unlabeled logs; his AAAI 2023 paper used reinforcement learning with a custom reward design.
Meet Karan ↗Varun Sahni
Building developer workflows around a detect, fix, and verify loop. His work on Code Buddy supports 12,000+ engineers—a practical foundation for keeping humans in the learning loop.
Meet Varun ↗Built for the way software changes.
Why build this now?
Teams are shipping more AI-written code. The code and its logs can change faster than static models or hand-written parsers keep up. Logmorph is being built around that distribution shift, rather than assuming yesterday's logs describe tomorrow's behavior.
What can I use today?
The investigation workspace is available now: import two log files, compare patterns, inspect source lines, search all logs, record review decisions, and save or export your work. The in-cloud model runtime, continuous training, incident replay gates, and rollback system are in development. The current workspace uses deterministic comparison, not model inference.
Does the current workspace upload my logs?
No. File analysis runs in your browser. Investigations are saved to this browser only when you choose Save locally. Backups include source logs and let you move an investigation to another browser. There is no account or cloud sync.
Who is building Logmorph?
Karan Bhukar and Varun Sahni. Karan's NYU research focuses on distribution shift and out-of-distribution coverage, building on BERTOps and reinforcement-learning research. Varun's work on Code Buddy applies a detect, fix, and verify loop to developer workflows. Meet Karan · Meet Varun