DwaraOne is the applied form of two published papers. Data Engineering-Driven World Models for Consequence-Aware Agentic Systems describes how governed data becomes the world model an agent reasons over — here, that is the preflight layer simulating leave requests against roster, crew coverage, blackout windows, and balance before execution. Data Engineering-Driven Hierarchical Planning for Consequence-Aware Agentic Systems describes decomposing intent into governed sub-plans — here, that is the life-event journey and the preview → confirm → submit → approve → audit lifecycle every transaction runs. The gold views on Databricks are the engineered substrate both papers assume. And the loop closes on itself: the hash-chained audit log records every request as a state, every decision as an action, and every transition as an outcome — exactly the state–action–outcome history the world-models research requires as training data, meaning DwaraOne generates the substrate for its own future learned intelligence as a byproduct of governance. Every conversation is also an observation: DwaraOne logs its own resolutions, retrieval misses, and corrections into the same governed ledger — the first loop of a governed self-improvement system, where the agent’s quality becomes a measurable gold view rather than a black box.
Vol. 01·Databricks Agentic Architecture
Governed onDatabricks · Unity Catalog
How DwaraOne
is composed.
A Databricks-powered conversational HR front door for the agentic enterprise. Context, tools, transactional state and memory — arranged as one governed loop.
§ 01 · The loop
- 31
- governed agent tools
- 8
- full policy documents
- 6
- lifecycle stages per transaction
- SHA-256
- chained audit entries
Live governance · session-scoped
Open reasoning trace - 0
- Turns this session
- 0
- Tool calls
- 0
- UC tables touched
- 31 · 8
- Tools registered · policies loaded
§ 00
What each one actually does hereThe pieces
- Unity Catalog
- Every read Agent Brahma makes resolves to a governed gold view — employee_360, leave_eligibility, training_due — and every answer cites its source on the card.
- Genie Spaces
- Three configured spaces stand behind natural-language analytics, so leadership questions become governed queries rather than exports.
- Policy Knowledge
- Eight full policy documents joined with employee_360 at question-time. Tenure, base, and dependents change the answer: six years of service means 24 leave days, not 22.
- Agent Tools
- Task Me Anything — every tool is a real write against Databricks, hash-chained into the audit ledger. 31 governed tools. Every write is a two-step intent — preview, then explicit confirmation — never a direct mutation. Preflight checks simulate consequences against roster, blackout windows, and crew coverage before anything is proposed.
- Lakebase
- Confirmed submissions insert into airhr_ops.requests; every state change appends to a SHA-256 hash-chained audit_log; a merge job carries the ledger into Delta under the same governance.
- Agent Memory
- Persona-scoped transcripts persist per role, and tool results ride along in conversation history — Agent Brahma remembers what his own tools told him a turn ago.
§ 01
Surfaces outside · agent + substrate insideThe trust boundary
Layered architecture · closed loopRead top → bottom
Surfaces · outside boundary
Employee
Manager
Org Head
↓
Trust boundary · Unity Catalog
Context
- Genie Spaces
- HR Ontology
- Policy Knowledge Base
↓
Agent
Agent Brahma — DwaraOne Concierge
readspreviewssubmitsapprovalspolicynext-best-actions
↓
Lifecycle
- 01 · preview
- 02 · confirm
- 03 · submit
- 04 · approve
- 05 · audit
↓
Substrate · Databricks
- Gold Viewsemployee_360 · leave_eligibility · training_due
- Lakebaseairhr_ops: requests · approvals · audit_log
- Delta mergestate → gold materialization
- Unity Cataloglineage · policies · access
↺ hash-chained audit · SHA-256
Every arrow crosses a governed boundary — Unity Catalog tables, Mosaic AI Agent Tools with audit flags, and a Lakebase-ready transaction ledger. Nothing bypasses lineage.
§ 02
Ready · Partial · Planned — no hand-wavingCapabilities
*
Demo transparency
Today's demo runs on seeded Gold views and an in-memory request ledger. The Lakebase persistence, live Mosaic AI tool registration, and per-user memory store are the production gaps listed above — each is a scoped follow-on, not a rewrite.§ 02.5
The substrate — what each primitive powers in DwaraOneDatabricks Free Edition
Everything on this page runs on the Free Edition today. Nothing below is aspirational — each primitive is wired to a specific DwaraOne surface, and the paid-tier delta is SLA and private networking, not capability.
- 01Unity Catalog
- Row- and column-level governance on every Gold view Agent Brahma reads — including the masked attrition_heatmap for Org Heads.How · A single catalog (airhr_one) owns schema, grants, masking functions, and lineage. Every server route resolves through it.Why · One place to prove who can see what — the answer to every audit question is already recorded.
- 02Delta Lake · Gold views
- airhr_one.gold.employee_360, team_readiness, experience_summary, policy_catalog — the substrate Agent Brahma reasons over.How · Silver → Gold jobs materialize agent-shaped views (joined, denormalized, typed) so tools never assemble facts at question-time.Why · Consistent answers across Employee, Manager, and Org Head — same view, different lens.
- 03SQL Warehouse (Serverless)
- Every server route under /api/airhr/* — profile, manager readiness, org metrics, policy check.How · Parameterized statements via the Statements API, warehouse 94af1bc5b884becc, 30-second wait timeout, results mapped into typed adapters.Why · No always-on cluster, no leaked credentials in the browser — reads scale to zero between requests.
- 04Genie Spaces
- Three role-scoped spaces — Employee, Manager, Org Head — behind natural-language Q&A.How · Each space is pinned to the Gold views its persona is allowed to see; Agent Brahma routes the question to the space that matches the caller's role.Why · Leadership analytics without exports — the query is governed the same way the dashboard would be.
- 05Vector Search
- Policy retrieval behind the Policy Navigator and Agent Brahma's get_policy_document tool.How · Policy chunks indexed with metadata (region, band, dependents); a tokenized keyword fallback covers cold-start and out-of-vocabulary queries.Why · Answers are grounded in the exact clause — with a citation on the card, not a hallucinated paraphrase.
- 06Workflows · Jobs
- Scheduled Gold refresh, telemetry rollups into agent_health, and periodic audit-log compaction.How · Jobs API orchestrates Silver→Gold, ledger→Delta merges, and eval-set replays; all runs emit lineage events into system tables.Why · The substrate keeps itself fresh — no cron on a laptop, no snapshot drift between demos.
- 07Lakebase (Postgres OLTP)
- The request ledger, consent ledger, and the SHA-256 hash-chained audit spine.How · Confirmed submissions insert into airhr_ops.requests; every state change appends to audit_log; a merge job carries both into Delta under the same catalog.Why · OLTP-grade writes and analytics-grade history share one governance boundary — no ETL detour.
- 08MLflow · Model Serving
- Reasoning traces, prompt versions, and Agent Brahma's eval sets — surfaced on the Org Head agent-health panel.How · Every conversation logs prompt version, tool sequence, retrieval hits, and user corrections as MLflow runs; regressions are caught before the next release.Why · The agent becomes a measured artifact, not a black box — quality is a gold view.
- 09AI Gateway
- Single governed egress for language and voice — Gemini for reasoning, ElevenLabs for speech — from server code only.How · Requests are proxied with per-tenant keys, PII redaction rules, and cost caps; browser code never sees a provider secret.Why · One place to switch models, cap spend, and prove nothing sensitive left the boundary.
- 10Lineage · System Tables
- The why-this trace, SLA countdowns on request cards, and the agent-health data source.How · system.access, system.query, and lineage tables are read the same way as any Gold view — so provenance is queryable, not folkloric.Why · Every decision Agent Brahma renders can be reconstructed from data Databricks already keeps.
The full agentic loop — reason, ground, act, audit — closes inside this Free Edition footprint. Moving to a paid tier changes SLAs, private networking, and warehouse size; it does not change the capability surface above.
§ 03
Applied form of two published papersResearch foundations
"One door. Every HR journey. Governed on Databricks." — Brahma Reddy