From video to usable APIs
Harsha co-founded Y Combinator-backed Moonshine. Its video-understanding APIs were used by thousands of developers around the world, bringing multimodal retrieval into tools people could build on.
Intelligence,
studied as a system.
We build AI research systems and the infrastructure that makes their results inspectable. Our work connects multimodal models, reliable execution, and experiments that can be reproduced and challenged.
Experimental systemsFigure 01
Allocated Completed Evidence received
Intelligent models
Multimodal learning & efficient inference
Experimental systems
Automated research & connected hardware
Computing infrastructure
Parallel execution & resource optimization
Reliable software
Traceability, evaluation & assurance
01 The lab
The model is one part of the experiment.
We study the system around it.
Meshia Labs brings applied AI research and systems engineering into the same workflow. We build our own compute and experiment infrastructure, connecting model behavior to resource use, software reliability, and real-world constraints. Our founder's background spans production autonomy, video-understanding APIs, and commercial vision-language systems.
Production experience
Research informed by systems people use.
Meshia is led by Harsha Gundala. The work below connects his prior experience with the research software we build today.
Production AI engineering
At Shield AI, Harsha contributed to production autonomy for V-BAT aircraft. Meshia's commercial VLM partnership with Rumi Labs is in production. That experience informs how we evaluate model behavior, integration constraints, and reliable execution.
Research software in use
Meshia's commercial research workspace is used by researchers from UC Berkeley, Stanford, and MIT to scale their research. We build around the practical work of running experiments, inspecting results, and retaining the evidence behind a conclusion.
Experience and adoption described by our founder, September 2026. University names identify researcher affiliations, not institutional contracts or endorsement. Prior-employer work is professional experience, not Meshia-owned technology.
02 Selected work
Questions we investigate.
Research directions in scientific computing, model behavior, and efficient inference. The diagrams below are illustrative, not measured benchmarks. Each question calls for a declared dataset, baseline, and reproducible evaluation.
RN-01 · STRUCTURAL BIOLOGY
Scaling structure prediction across a compute fleet
Parallel structure prediction creates a second problem: deciding which results deserve further validation. We investigate calibration, failure cases, and the compute cost of that decision.
Research question How should prediction confidence guide validation when structure inference runs in parallel?
Calibration evaluation focus
Cost per result resource measure
RN-02 · MODEL SAFETY
Selective unlearning and retained model capability
An unlearning method must remove the intended information while preserving useful behavior elsewhere. We study target and control performance together, including refusals and unintended regressions.
Research question Can a model selectively forget a target domain while preserving useful behavior elsewhere?
Target & control paired evaluation
Regressions explicitly tested
RN-03 · EFFICIENT INFERENCE
The quality–cost frontier of on-device inference
Our production VLM experience informs research on distillation, quantization, and architecture choices. We evaluate quality alongside device latency, memory, and energy under matched test conditions.
Research question Which combinations of model compression and hardware preserve quality under a constrained compute budget?
Matched baselines comparison method
Device profiling evidence required
03 Research areas
A connected research agenda.
Six areas of inquiry, linked by a common interest in how intelligent systems behave beyond the model itself.
R-01
Multimodal AI & machine learning
Vision-language and LLM fine-tuning, contrastive pretraining, and multimodal fusion for perception, monitoring, and alerting. Open weights in, mission-tuned checkpoints out, with repeatable evaluation at every step.
R-02
Autonomous & multi-agent systems
Research workflows that coordinate agents, preserve intermediate results, and expose uncertain actions for review. We study recovery, verification, and the boundaries of useful human oversight.
R-03
High-performance computing
Scheduling, placement, and fault tolerance for large workloads on heterogeneous compute, with measured placement and recovery behavior.
R-04
Software modernization & assurance
Legacy system refactoring, software debloating, and provenance-aware build pipelines that make large codebases smaller, safer, auditable.
R-05
Model efficiency & edge hardware
Distillation, quantization, and hardware-software co-design that put capable models on devices in disconnected, contested environments.
R-06
Data, knowledge & novel signals
Source-linked knowledge graphs, hybrid retrieval, and detection of novel signals in noisy multimodal data, grounded in versioned, auditable corpora so every inference can be traced to its source.
04 Method
The infrastructure is part of the method.
A useful experiment needs more than a result. It needs a declared comparison, a repeatable execution path, and enough context for another researcher to examine what happened.
01
Define the comparison
Specify the hypothesis, baseline, evaluation data, and resource budget before interpreting a result.
02
Run controlled variations
Distribute seeds, ablations, and parameter sweeps across available compute while retaining the conditions of each run.
03
Preserve the evidence
Keep inputs, configuration, outputs, and evaluation artifacts together so a conclusion can be traced to its source.
04
Test the boundary
Investigate failure cases, repeat the comparison under changed conditions, and separate measured behavior from the next hypothesis.
Experiment throughputConceptual model
05 Research partnerships
Research shaped around a consequential question.
We pursue collaborative and sponsored research in applied AI, computing, and autonomous systems, including SBIR/STTR and dual-use R&D. We are interested in problems with a clear technical uncertainty and a practical path to evaluation.
Sponsored research
We prepare focused research proposals with explicit hypotheses, work packages, acceptance tests, and a transition plan. We distinguish existing capability from work the proposed funding would make possible.
Feasibility studies
Our commercial workspace provides an engineering foundation. Each new research claim still needs its own evidence: a competent baseline, a reproducible test, and an honest account of where the method fails.
Deployment realism
We define the data rights, integration interfaces, available expertise, and deployment boundary before promising a pilot. A local prototype is not a security accreditation or proof of operational readiness.
Let's investigate
a difficult problem.
We welcome conversations with researchers, program managers, and engineering teams. Tell us about the question, the constraints, and what a useful result would look like.