VALESKA
A Finding Must Carry Its Receipt
September 8 notes: a failed memory trial drove repairs to how Valeska preserves findings and their evidence. Later verification showed qualified reuse, not universal adoption.
Read articleAVANTI AI INNOVATORS / DAVID GARGAN
Ideas into practice.
Notes on AI, integration, and building useful systems.
VALESKA
The architecture of governed knowledge: retrieval, provenance, permissions, and institutional memory in VALESKA.
Multi-agent Systems
DaveRouter and Commons: execution, shared decisions, work ownership, and coordination across multi-agent environments.
StoryTeller
The technical architecture of visual storytelling: story and scene models, generation workflows, and the connections between narrative and visual assets.
VALESKA
Useful Memory Needs Boundaries, and Permissions Need a Source
I test whether Valeska remembers useful decisions rather than machine noise, and connect document permissions to their source without mistaking that connection for a security guarantee.
Read article
VALESKA
Ship Without the ECM, and Without Phoning Home
A clean-host test exposes the difference between a standalone AI package and one that merely starts. I remove hidden dependencies, test offline operation, and keep the limits visible.
Read article
VALESKA
Survey the Work Before You Ingest It
Research notes for August 21: an internal survey finds fourteen discovery defects, distorted file counts, and reasons to settle document scope before collection.
Read article
VALESKA
Retrieval Must Be Scored, Not Felt
Research notes for August 20: a 33-question retrieval test improves coverage but not the first answer, while a separate timing correction exposes a 48-fold measurement error.
Read article
VALESKA
Classification Must Be Saved, and Records Must Be Findable
A classifier finally saves its result, while a records-management test exposes the gap between applying a label, enforcing retention, and delivering trustworthy search.
Read article
VALESKA
The Corpus Must Be Hygienic Before It Is Smart
August 18 research notes: measured provenance checks correct a storage assumption, while a copied collection falls from 7,574 indexed chunks to 49 without proving better answers.
Read article
VALESKA
A Document Has to Become a Knowledge Unit, Not a Blob
August 16 research notes: section-based knowledge units and an internally restricted source selection give VALESKA a more inspectable foundation, not proof of training readiness.
Read article
VALESKA
Ingest Must Read the Formats People Actually Send
August 11 research notes: broader document extraction, limited PDF and OCR evidence, and collaboration checks that must disclose when they are not protecting work.
Read article
VALESKA
Memory Must Serve a Live Agent, Not a Demo Session
August 7 research notes: a narrow connection for email work, tested memory operations, an unfinished retry connection, and a correction for accented text.
Read article
VALESKA
Captions Have to Be the Search Text, Not a Sidecar Demo
Today I pivot captions from a CLI prototype to an ingest-time pipeline and API, and I measure a gold set that says bare labels still beat my hierarchical search.
Read article
VALESKA
A Relationship Is Not Evidence If Nobody Could See It
Today I require ACL-visible evidence for ontology relationships and carry that provenance into exported training datasets.
Read article