Obseria Research: the science behind reliable systems
We research how machines can explain software failures with evidence, store telemetry efficiently and keep observability open and private by design.
What Obseria Research works on
Our research questions come from real operational problems and end up in the product.Causal inference for root cause
Using topology and change events to separate causes from symptoms instead of ranking correlated metrics.
Anomaly detection and forecasting
Seasonality-aware baselines and forecasts that adapt to changing workloads and keep alert noise low.
Agentic investigations
AI agents that query telemetry step by step and document every conclusion in a reviewable chain of evidence.
Efficient telemetry storage
Columnar layouts, compression and schema-on-read techniques for an index-free lakehouse at petabyte scale.
Sampling
Tail-based and adaptive sampling strategies that keep the traces that matter while controlling volume and cost.
LLM evaluation
Methods to measure quality, groundedness and safety of LLM and agent outputs in production.
Our mission: answers you can verify
Obseria Research exists to make complex software systems understandable. We focus on methods that explain why something happened, not only that it happened. Every result has to hold up in production conditions: noisy data, partial instrumentation and constant change.
- Problems sourced from real incidents and operator workflows
- Results measured against reproducible evaluation setups
- Explainable outputs that engineers can check and challenge
Open source contributions to OpenTelemetry, Prometheus and Perses
Observability should be built on open standards. Our engineers contribute to OpenTelemetry, Prometheus and Perses, including semantic conventions, Collector components, specification discussions and dashboards-as-code tooling. What we learn in research flows back into these communities.
- OpenTelemetry: semantic conventions, Collector and SDK work
- Prometheus: compatibility and query engine topics
- Perses: open dashboards-as-code specification
- OTLP in and out, so your data is never locked in
Research principles: evidence, explainability, privacy
We prefer evidence over guesses: a root-cause claim must point to the data that supports it. Models must be explainable enough for an on-call engineer to trust or reject them. And privacy is non-negotiable: customer data is never used to train AI models.
- Evidence over guesses: every conclusion links to its data
- Explainability over black boxes
- Customer data is never used for training
- PII masking and field-level access apply to research tooling too
From research to product, together with universities
Research at Obseria is organised in small teams that work directly with product engineering, so prototypes become features rather than slide decks. Through the University Alliance we work with teaching staff and students on observability curricula. If these problems excite you, we are hiring.
- Shared roadmaps between research and product teams
- Results published on the engineering blog
- University Alliance for teaching and student projects
- Open roles in research, data and platform engineering
How research reaches the product
- Step 1
Observe the problem
We start from real operational pain: noisy alerts, slow root-cause analysis, rising telemetry cost.
- Step 2
Prototype and evaluate
Ideas are tested against reproducible evaluation setups before they reach any customer.
- Step 3
Ship behind evidence
Features ship in Obseria Intelligence, Pipeline or Lake with explanations and controls built in.
- Step 4
Share and contribute
Learnings are published on the blog and contributed back to open source where they fit.
Latest from the engineering blog
Migrating from proprietary agents to OpenTelemetry: step-by-step
Read articleOpenTelemetry Collector in Production: Best Practices for SREs
Read articleTail-Based Sampling in the OpenTelemetry Collector: Keep Every Error, Save 80% of the Cost
Read articleFrequently asked questions
Causal inference for root-cause analysis, anomaly detection and forecasting, agentic investigations with a chain of evidence, efficient telemetry storage and compression, sampling and LLM evaluation.
No. Customer data is never used to train AI models. Research relies on synthetic, public or internally generated data.
Yes. Our engineers contribute to OpenTelemetry, Prometheus and Perses, and Obseria supports OTLP ingestion and export so you are never locked in.
Research teams work directly with product engineering. Results are evaluated reproducibly and then shipped in Obseria Intelligence, Pipeline or Lake with explanations built in.
Yes. The University Alliance supports teaching staff and students, and you can reach the research team via the contact page or apply for open roles on the careers page.
Help us make systems explainable
Follow our engineering blog or join the team working on causal AI and open telemetry.