Matt Redmond

Matt Redmond contributed to Hex’s failure-funnel framework, which breaks AI evaluations into stages to identify where a system fails. In April 2025, Bryan Bischof credited Redmond with co-developing it alongside Charlene Chambliss, Izzy Miller and Sajel Shah. Redmond also publishes model-quantization work and maintains a repository for learning representations of relational databases.[4][6][7]

The framework replaces a single pass/fail result with ordered, cascading checks. Hex’s SQL example asks whether retrieval supplied the necessary tables, whether the agent invented or selected incorrect tables and columns, whether the query executed, and whether it returned correct data. Each upstream failure masks later stages: a query containing a nonexistent column cannot proceed successfully through execution to correct results. This ordering helps distinguish a root failure from its downstream symptoms.[5]

Izzy Miller’s April 2025 explanation recounts why that distinction mattered to the team. Apparent table and column hallucinations had prompted unsuccessful attempts to strengthen instructions. The team eventually traced them to retrieval bugs that left the model without the needed context. Fixing retrieval improved performance.[5]

The same approach exposed a specific GPT-4.1 behavior: lowercasing and quoting column identifiers caused database errors. Miller reported that a one-line prompt fix yielded 12 additional correct-data passes on Hex’s private evaluations. Alongside the funnel, the framework uses “flux” to track how far individual evaluations move between stages and orbit charts to show improvements and regressions, with semantically similar cases clustered together. These tools describe the shared framework that Bischof credited Redmond with co-developing.[5][4]

Redmond’s four-bit LLaDA model card records a concrete hardware constraint: “Had to quantize this so it would fit on a 3080TI - all I've got!” He used Modal for the quantization and published a reproduction script. It loads the upstream GSAI-ML/LLaDA-8B-Instruct model and configures four-bit GPTQ quantization with a group size of 128. His Hugging Face profile lists both four-bit and eight-bit GPTQ artifacts, each displaying an update date of June 26, 2025.[6][2]

Redmond’s Tributary repository describes a foundation-model system intended to learn internal representations of relational databases through prediction tasks. It explicitly credits Rishabh Ranjan’s Relational Transformer for the predominant core ideas, calling Tributary “mostly a small re-implementation plus some slight engineer-y modifications.”[7]

Tributary represents relationships between database rows as a graph and trains with a masked language model objective against prediction task tables. Its targets include numeric, timestamp, boolean and categorical values, with different losses for those types. Stated modifications include treating null-versus-present values as learning signals and adding categorical and identifier semantic types. Identifier values themselves are omitted; their presence or absence is retained as potentially useful signal.[7]

Redmond uses mrdmnd on GitHub and Hugging Face; his GitHub profile names him Matthew Redmond and describes him as a data scientist and computational geometer interested in hardware performance. As of September 20, 2026, that profile listed @hex-inc and San Francisco. Latent Space’s June 2025 AI Engineer World’s Fair recap acknowledged him through Bryan Bischof in its Architects 2 track-host credits.[3][2][1]