Performance changes across data slices, prompts, environments, or time and the cause is unclear.
Independent AI research & engineering · New Jersey, USA
AI systems,
made legible.
Topologica LLC is an independent AI research and engineering lab based in New Jersey. We help organizations understand, evaluate, and improve machine-learning systems when model behavior is difficult to diagnose, data cannot or should not be centralized, or a research prototype has to work under real operating constraints.
Model diagnostics
Federated ML
AI systems
Research engineering
EvaluationFailure modes, calibration, drift, representations
AdaptationFine-tuning, LoRA/PEFT, held-out evaluation
FederationDistributed training and evaluation workflows
SystemsAgents, automation, research software
Engagement fit
Useful when the next technical move is not obvious.
Topologica is most useful when a system already matters enough that a vague answer is expensive. Common starting points include diagnosis, constrained data, research-to-production handoff, and cross-system AI workflows.
Centralizing data is undesirable, restricted, or simply unnecessary for the question being asked.
A prototype or analysis needs tests, reproducibility, interfaces, and a defensible technical handoff.
Agents, APIs, retrieval, or automation cross system boundaries and need explicit state and failure handling.
Capabilities
Work scoped around the actual decision.
Topologica works across machine learning, software, and technical product systems. Engagements start from the question, constraints, and evidence available, then use the smallest approach that can resolve the problem.
Model evaluation & diagnostics
Evaluation design and failure analysis across predictions, embeddings, activations, calibration, drift, data slices, representation geometry, and model behavior.
Model adaptation
Fine-tuning and parameter-efficient adaptation, including LoRA and PEFT workflows, with dataset hygiene, held-out evaluation, reproducible training artifacts, and explicit limitations.
Federated & distributed ML
Distributed training and evaluation for settings where centralizing data is undesirable or constrained, including systems developed with Flower and other federated-learning frameworks.
AI systems & automation
Agent workflows, APIs, MCP interfaces, retrieval, orchestration, and human-in-the-loop automation designed around observable state, failure handling, and maintainable interfaces.
Research engineering
Turn exploratory technical work into tested code, manifests, reproducible evaluations, documentation, and handoff packages that another technical team can inspect.
AI system audits & architecture
Independent review of AI products and pipelines: architecture, dependencies, evaluation design, observability, provenance, deployment assumptions, and operational failure modes.
Deployment & distributed systems
When data, compute, or trust boundaries matter.
Topologica works across local, private-cloud, managed-cloud, and federated architectures when centralized processing is not the right default. Work can include Flower-based simulation, distributed training or evaluation, aggregation, and integration with existing software, with data movement and trust boundaries made explicit.
Federated training & evaluation
Design and implementation of federated training, evaluation, simulation, and aggregation workflows using Flower and related frameworks.
Deployment architecture
Local, private-cloud, managed-cloud, or federated execution can be selected according to data location, trust boundaries, latency, and operational constraints.
Privacy and security are separate claims
Threat model, privacy mechanism, deployment topology, and operating controls are evaluated explicitly. Federated architecture alone is not a security or compliance certification.
References to Flower or other frameworks describe technologies Topologica can integrate and do not imply affiliation, endorsement, certification, or a particular regulatory status.
Working method
Scope the decision. Then test what matters.
Complex AI work becomes easier to reason about when the decision, evidence boundary, and failure conditions are explicit. The objective is not more machinery. It is a better-supported next action.
- 01
Define the decision
Specify what must be learned, changed, or built, what constraints matter, and what would count as sufficient evidence.
- 02
Inspect the system
Examine data generation, model behavior, software boundaries, interfaces, dependencies, and operational context before selecting an intervention.
- 03
Run discriminating tests
Use baselines, held-out evaluation, failure slices, sensitivity checks, reproducible runs, and counterexamples where they resolve the actual uncertainty.
- 04
Deliver inspectable work
Provide the implementation, evidence, limitations, and next actions in a form another technical team can review and continue.
Selected public work
Selected work, with a public record.
Public repositories and research artifacts show how selected work is built and documented. They are examples, not client endorsements, production guarantees, or a complete inventory of Topologica's work.
A / MACHINE LEARNING
TopoGeoML
Open-source topology-aware machine-learning software and a preregistered graph-classification study with matched baselines and reproducible computational experiments.
B / RESEARCH INFRASTRUCTURE
Semantic Scholar MCP
Open-source MCP infrastructure for scholarly search, citation-graph traversal, author profiles, recommendations, and structured machine-consumable research workflows.
C / PUBLIC RECORD
Technical identity
Selected public profiles provide direct access to code, research identity, competitive machine learning work, and professional background.
About
Independent by design.
Topologica LLC is an independent AI research and engineering company based in New Jersey. Founded by Santiago Maniches, it works across machine learning, data systems, scientific computing, software engineering, and technical product development. Public repositories and research are a selective record, not a complete inventory of the lab's work.
Contact
Start with the technical problem.
For consulting, technical review, applied research, or implementation inquiries. A useful first message is the system, what is failing or uncertain, and the decision you need to make. For sensitive or regulated information, establish an appropriate handling channel before transferring data.