Evidence, not decoration
This page separates approved customer evidence, public technical artifacts, founder-confirmed operating experience, research, and work still being prepared. Missing proof remains visibly missing.
These artifacts make the proposed architecture discipline inspectable. They are not customer case studies, measured customer outcomes, certifications, or proof that every control has been implemented in a production environment.
Divides business intent, agent behavior, execution, context, security policy, evaluation, runtime, release, and rollback across accountable owners.
Connects enterprise AI access, coding-agent execution, and workflow-agent action to identity, data/IP, model, MCP/tool, sandbox, and assurance controls.
Shows how model access, tool execution, evidence, data classification, provider terms, cost, and support ownership become a reviewable decision.
Customer evidence: no enterprise architecture, agent-operations, or secure-enablement customer case study is currently approved for public use. We publish one honest preparation note rather than placeholder case-study cards.
This is product-position evidence, not a claim of a completed customer deployment. Each item is reviewed against the current early-access product destination.
Define and configure a workflow for the operating lifecycle; this is not presented as a general-purpose agent builder.
Lifecycle support for testing, deployment, operation, review, exceptions, and improvement according to the current product release.
Review, approval, and intervention patterns for workflows whose risk warrants human oversight.
An operating principle centered on exportability and reduced platform dependence, subject to the current supported formats.
Zentash is early access. “Create” means defining or configuring a workflow for its operating lifecycle; it is not presented as a general-purpose agent builder or as automatic IAM, network, DLP, MCP-isolation, or runtime-security enforcement.
This lane preserves the existing language-model run inventory, OpenLanguageModel attribution, publications, patent records, and clear scaling limitations.
Founder-confirmed broader internal count. The itemized, publishable inventory is being curated.
OpenLanguageModel’s public materials describe 30+ language-model runs from 100M to 1B scale.
We have not yet delivered one. Larger work starts with scaled experiments and explicit gates.
Records will include parameter count, objective, dataset summary, result summary, artifact link, and public/summary-only status. Only approved entries will appear.
We are reconciling internal experiment records with artifact availability and publication approval. We will not create placeholder cards for unpublished runs.
An MIT-licensed, PyTorch-native library for building, training, teaching, and researching transformer language models.
Tavish Mankash is one of three named software authors, alongside Vardhaman Kalloli and Keshava Prasad. We describe the attribution precisely rather than presenting OLM as a project exclusive to Mankash AI.
Research on extracting medication names and dosages from simulated Indian medical-record images using fine-tuned multimodal language models.
View on arXivThese patents are cited as individual technical background. They are assigned to Medyug Technology Private Ltd; the website does not claim ownership by Mankash AI.
A reserved publication module will document problem and dataset provenance, baselines, metric definitions, hardware, training configuration, quality/cost/latency results, failure analysis, artifacts, and limitations.
Architecture consultation
Bring the portfolio, security boundary, agent lifecycle, or model decision. We will identify the narrowest useful first engagement.