Systems
We design and build AI systems for specific organizational contexts. Depending on the problem, this may involve retrieval over private information, document and data processing, evaluation infrastructure, model adaptation, workflow automation, or systems that take actions under defined controls.
A system is not complete when a model produces an impressive demonstration. It is complete when its performance can be measured, its limitations are understood, its operation is documented, and the people responsible for it can run it reliably.
Applied system
Novalis
Novalis answers questions about an organization's own documents and links each substantive claim to the passage that supports it.
Organizations often hold critical knowledge in policies, reports, correspondence, scans, spreadsheets, and other document collections that are difficult to search consistently. Novalis makes those collections available through an ordinary-language interface while preserving a direct connection between each answer and its source.
Where the available documents do not support an answer, Novalis says so. It does not present an unsupported conclusion merely because one is linguistically plausible.
Capabilities
- Works with your existing documents
- Novalis processes office files, PDFs, scans, spreadsheets, correspondence, and other common formats.
- Provides traceable answers
- Each claim is linked to the source passage from which it was derived.
- Expresses uncertainty explicitly
- When the document collection does not support an answer, the system declines to provide one.
- Operates within your infrastructure
- Novalis runs in your environment and makes no external calls by default.
- Supports access control
- Collections can be separated by team, subject, or other organizational boundaries, with access governed at the individual level.
- Includes an operator console
- Your team can administer the system, manage collections, and monitor operations without requiring shell access.
Other areas of work
Novalis is one applied system developed by the lab. Our broader work includes:
- Retrieval and question answering over private data
- Document extraction and classification
- Evaluation, benchmarking, and quality measurement
- Model selection, adaptation, and fine-tuning
- Workflow automation and structured decision support
- Agentic systems with bounded and auditable actions
- AI systems that must operate under privacy, security, or regulatory constraints
- Production deployment, monitoring, and operational handover
We select the technical approach according to the problem. We do not assume that every problem requires a larger model, a retrieval system, or an autonomous agent.
Engagements
Our engagements cover the full path from problem definition to reliable operation. The exact sequence varies by project, but the work generally proceeds through the following stages.
Discovery and problem definition
We begin by clarifying the decision, workflow, or operational problem the system is expected to address. This includes understanding the users, existing processes, available data, technical environment, security requirements, and unacceptable failure modes.
The result is a defined problem rather than a general request to use AI. We identify what should be automated, what should remain under human control, and what evidence would demonstrate that the system is useful.
Feasibility and risk assessment
We assess whether the available data, models, and operating environment are sufficient for the proposed objective. Where appropriate, we examine data quality, retrieval performance, model behavior, integration constraints, privacy requirements, and likely sources of error.
We also identify cases in which the proposed approach is unlikely to provide sufficient value. If the problem is not a good fit for an AI system, we say so before significant implementation begins.
Evaluation design
Before building toward production, we define how performance will be assessed. This may include representative examples, expert-reviewed test sets, task-specific quality criteria, refusal and uncertainty tests, latency and cost requirements, and operational acceptance criteria.
The evaluation is developed against the intended use case rather than relying solely on public benchmarks. This gives both parties a shared basis for deciding whether the system is improving and whether it is ready to proceed.
Prototype and technical validation
We build a focused prototype using representative data and the intended technical environment wherever possible. The purpose is not to create a polished demonstration, but to test the central assumptions of the proposed system.
At this stage we compare technical approaches, expose failure modes, and determine which parts of the workflow can be handled reliably by software and which require human review.
System development
Once the approach has been validated, we develop the system around the operational requirements of the organization. This may include data pipelines, retrieval, model orchestration, interfaces, access controls, evaluation tooling, audit logs, and integrations with existing systems.
Working software is delivered early so that it can be tested by the people who will use and operate it.
Deployment and integration
The system is deployed within the organization's infrastructure and subjected to the relevant security, privacy, and operational reviews. We test it under realistic conditions, including incomplete data, ambiguous requests, unexpected inputs, and changes to the underlying information.
Deployment is treated as an engineering and organizational process, not as the final step after the AI work is finished.
Training and handover
We train the responsible team to operate the system, review outputs, manage data, investigate failures, and respond to routine issues. Documentation covers the system's architecture, assumptions, evaluation results, limitations, and operating procedures.
A deployment is not complete until the client can operate it without depending on informal knowledge held by the implementation team.
Continued improvement
After handover, we can remain involved through model upgrades, extensions, evaluation updates, monitoring, and investigation of production failures. The objective is to improve the system in response to observed use rather than to introduce change for its own sake.
Where ongoing support is not required, the system is designed and documented to remain maintainable by the client's team.
To discuss an engagement, write to hello@mnemiclabs.com.