ARA

ARA is the decision data plane built by ARA AI Labs, a DPIIT-recognized startup based in Pune, India. It ships as a single compiled binary that runs synchronously in the ML serving hot path, binding entity state, feature vectors, model versions, and decision outputs into a permanent, append-only record at the exact microsecond of inference. The product targets production ML teams in regulated industries—credit, fraud, KYC, insurance—who need to answer "why did the model decide that?" months after the fact, without reconstructing from logs.

What It Is

ARA sits between your feature store and model serving layer. At inference time, the serving path calls ARA via a real-time API to register the entity, its full feature vector, and the decision output in a single write. The result is a temporally ordered, cryptographically chained snapshot per entity that is immutable after creation. Because every write is timestamped at microsecond precision and chained to the prior snapshot for the same entity, the record is tamper-evident by design—no instrumentation layer, no post-hoc enrichment, no reconstruction gap.

The platform supports two integration modes:

• Ledger mode: ARA sits alongside an existing feature store. Model serving calls ARA via API to register entity, features, and decision in one write.
• Native mode: ARA acts as the feature store itself. Training data is ingested once; features are served at low latency; decision records are captured automatically in the same plane.

Core Capabilities

• Exact replay: Retrieve any past decision by request ID and recover the exact feature vector the model saw at that microsecond via get_snapshot(as_of=decision_ts). The chain_valid field confirms cryptographic chain of custody.
• Point-in-time queries: get_snapshot(), get_latest(), get_feature_history(), and diff() all operate against the full temporal index—not a log reconstruction.
• Drift and skew monitoring: compute_drift() runs CUSUM accumulation against registered training statistics; compute_skew() flags individual values by z-score. Sentinel pair correlation tracking detects silent pipeline changes.
• Training data extraction: get_training_data() performs a point-in-time join against the temporal index, eliminating label leakage by returning only feature values that existed at or before each decision's timestamp.
• Right-to-erasure: forget_entity() suppresses all reads for an entity server-side, implementing GDPR Article 17.
• Replay Console: A Streamlit browser UI for timeline inspection, forensic drift analysis, two-point diffs, blast radius estimation, and entity comparison—without writing code.

Architecture and Deployment Model

ARA ships as a self-contained binary with no runtime dependencies. It requires Linux x86_64 or ARM64 (aarch64), 2 vCPU and 2 GB RAM minimum, and NVMe storage for lowest latency. A Docker image is published for linux/amd64. The server starts a worker pool sized to the machine, each worker listening on its own port starting at 50051. Client SDKs are available for Python (3.12+, via PyPI as ara-labs-sdk) and Java (21+, via Maven Central as ai.aralabs:ara-java-sdk). Both SDKs use a high-performance binary protocol (FlatBuffers over TCP) and are fully interoperable against the same server.

Community Edition runs as a single node. Enterprise Edition adds multi-node high-availability replication with automated failover, a 99.9% uptime SLA, dedicated engineering support, RBAC, SSO, and compliance tooling—all running on the customer's own infrastructure with no cloud dependency and no data leaving the environment.

Regulatory Positioning

ARA's architecture documentation maps the decision record explicitly to three regulatory frameworks:

• EU AI Act Article 13 (transparency for high-risk AI systems): enforcement deadline August 2, 2026 for credit, fraud, KYC, and insurance AI.
• SR 11-7 Model Risk Management: the snapshot record satisfies the documentation requirement with exact data, not post-hoc reconstruction.
• DORA Article 9 (digital resilience): replay capability enables post-incident analysis and demonstration of recovery to auditors.

Structured audit artifact exports (PDF and JSON) are planned for Enterprise, with scope driven by design-partner requirements.

Update: v1.0 Community Edition

The homepage identifies the current release as Decision Data Plane · v1.0: Community Edition. The GitHub issue tracker (created July 2026) references the latest release at aralabs.ai/download. SDK versions 1.0.1 are current for both the Python and Java SDKs. The core architecture is the subject of a U.S. non-provisional patent application filed in 2026. The company was founded by Tushar Haldar, formerly a Staff Software Engineer at Unity and previously at Oracle, who returned to India in 2025 to build ARA from the storage engine up.