Latent Alpha builds geometric AI for industries where the answer hides in the shape of the data. We translate complex, high-dimensional signal into solutions grounded in mathematics you can trust, ship, and explain.
Most modern data — markets, molecules, media, machinery — lives on a low-dimensional manifold inside an enormous ambient space. Standard models flatten that geometry and lose what matters. Latent Alpha doesn't.
We design models and tools that respect structure. The result is fewer parameters, better generalisation, and predictions a domain expert can interrogate.
Latent Alpha is a 2025 Yale spinout from the Krishnaswamy Lab — with over 30 AI/ML technologies and 50+ peer-reviewed publications, applied across industries including life sciences, banking & finance, media & advertising and commerce.
Biology is geometric. States, trajectories, and structure all live on a manifold. LLMs flatten that into tokens; geometric deep learning preserves the structure data actually occupy. Both are powerful — only one respects the geometry the data is written in.
Big problems — and most production ML — isn't really a language problem, although some try to make it. It's a structure problem. When the data lives on a manifold (cells, markets, molecules, sensor fields), geometric methods are smaller, faster, more accurate, and easier to defend and interrogate. Both families are powerful — only one respects the geometry the data is written in.
Every engagement assembles from the same core toolkit. We do a few things deeply rather than many things shallowly — and we publish what we learn.
Diffusion-geometric embeddings (PHATE, MAGIC, MIOFlow) for high-dimensional data — preserving local and global structure where t-SNE and UMAP fail.
Score-based and flow-matching models conditioned on geometry. Sample from data distributions with explicit, controllable trajectories.
Architectures that bake symmetry into the model — translation, rotation, permutation — so they generalise from a fraction of the data.
Domain-specific pretraining at scale, with the geometric inductive biases that let a single model serve many downstream tasks reliably.
Counterfactual modelling on observational data — interventions, treatment effects, and confounder discovery, grounded in structure.
Tools that let domain experts read what a model has actually learned — visualisations, attributions, and certified bounds. Models you can defend.
We partner with teams whose data is high-dimensional, noisy, or relational — and where being right matters more than being fast.
Returns, volatility, and cross-asset relationships sit on low-dimensional, curved structures that drift with regime. Geometric deep learning models that geometry directly — preserving local neighbourhoods, respecting symmetries, and exposing the topology of the risk surface rather than collapsing it to a few principal components.
The methods we publish were built for this domain: PHATE, MELD, MIOFlow, ImmunoStruct. Cells move along developmental manifolds; proteins fold in 3D; perturbations propagate through molecular graphs. Equivariant and manifold-aware models capture this without throwing away the structure that makes biology biology.
Identifier deprecation and platform fragmentation broke the lookalike era. Embedding audiences on a manifold — where similarity is geodesic, not categorical — produces targeting and lift estimates that survive cold start, sparse signal, and creative drift. Graph models handle attribution as a propagation problem, not a last-touch one.
Records, lineages, and schemas form graphs; entities live in embedding spaces; quality issues cluster by topology. We bring graph intelligence to the modern data stack — ingesting broadly from Postgres, Snowflake, BigQuery, Kafka, S3, and native graph stores (Neo4j, TigerGraph, Neptune) — and treat the platform itself as a geometric object instead of a pile of tables.
Generic deep nets ignore conservation laws and symmetries that energy and climate systems obey by definition. Graph and equivariant architectures — combined with physics-informed losses — produce forecasts and control policies that respect the dynamics, generalise across topologies, and stay stable under distribution shift.
Robots rotate, translate, and manipulate objects in 3D. Industrial processes have invariances and conservation laws that any sane model should bake in, not learn from scratch. Equivariant networks and Lie-group-aware policies need orders of magnitude less data and generalise across hardware, factories, and seasons.
Six months of returns for the S&P 500, embedded into three dimensions and coloured by sector. Hover any ticker; rotate, pan, and zoom.
The geometry shows what flat factor models can't — sectors cluster, regimes separate, and outliers sit where they should. Same idea, applied to your data: portfolios, audiences, cells, sensors, customers.
Selected work from our team and collaborators that defines the toolkit we apply across industries. The full list lives on arXiv and Scholar.
Latent Alpha spun out of the Krishnaswamy Lab at Yale. Our team authored the diffusion-geometry methods we now apply across industries.
Associate Professor of Computer Science and Genetics at Yale. PhD from University of Michigan EECS; postdoctoral training at Columbia systems biology. One of the leaders in geometric deep learning and graph neural network research.
LinkedIn
Over 30 years of life-science experience: banker, founder, inventor, senior management, investor and EIR at Yale. Expertise across therapeutic areas, machine learning, geometric deep learning, and building great teams.
LinkedIn
Twenty years managing technology, IP, and scientifically-based business alliances with biotech and pharma worldwide. Director of Business Development at Yale Ventures and EIR. Group Leader, Alliance Management & Biomarkers Discovery at CuraGen.
LinkedIn
15+ years of biotech VC expertise (SR One, Abingworth, Trekk Venture Partners). Public and private board experience with 18+ biotechnology companies, including Nimbus Therapeutics. Deep scientific and operational background.
LinkedInFor partnership, investment, or career inquiries, reach out directly. We respond to qualified outreach within one business week.
We respond to qualified outreach within one business week.
