The HPC Journey
HPC research lead Prof. Saraswati Iyer at IISc Bangalore on AIRAWAT taps a single button. Twelve stops happen in 120 milliseconds. Every HPC-specific regulator pack and domain plugin is enforced inline.
HPC research lead Prof. Saraswati Iyer:
“Submit weather-forecast ensemble: 200-member, 90-day rollout, 0.5° grid. Partition assignment? GPU vs CPU mix? Estimated wall-clock + cost? Compare to last 5 ensembles I ran.”
The 12 stops — tailored for HPC
Same skeleton as the cross-sector journey. Only the regulator pack, domain plugin, and validation ceremony change.
Arrival
Single ingress; HPC research lead's request signed at the gateway with sector context.
Identity
Prof. Saraswati Iyer's DID resolved with HPC-specific role tier and authorisation scope.
Sentinel pre-check
Universal Policy Ring evaluates against HPC-active regulators: Sovereign-research data licence + IMD data-use terms + iDEX classification (where applicable) + DPDP for any embedded location data.
Routing — "do we even need a model?"
Most HPC routine queries answer from neural mesh alone. Confidence threshold 0.85.
Compliance frame
Loads 12 of 421 controls that apply to THIS query: Sovereign-research data licence + IMD data-use terms + iDEX classification (where applicable) + DPDP for any embedded location data.
Memory recall — the differentiator
Retrieves what the institution has already decided. Last 5 similar ensembles by this researcher: average 14 hrs wall-clock on 200 H100 GPUs, 87% utilisation, optimal partition was 8x25 across nodes. Two ensembles hit memory pressure at hour 9; suggest larger node-RAM allocation for 200-member runs.
Knowledge graph traversal
SPARQL traversal of the institution's HPC ontology in adios-deeproot; connects scattered records into reasoning-ready structure.
Reasoning — SKIPPED if mesh confidence high
Routine HPC queries: skip. Complex synthesis: on-prem inference call (Sarvam, BharatGen, or sector-specific model). NEVER external API for sovereign-data queries.
Domain AI — package in HPC language
Slurm/PBS submit script with optimal partition + memory + GPU spec, pre-flight validation against AIRAWAT's current capacity, expected completion timestamp, billing forecast.
Validation — the hinge
single-key researcher validation; job submitted to Slurm; outcome metrics (actual wall-clock, utilisation) update pattern on completion.
Compounding — Patent S8
Validation becomes a permanent pattern in the institution's neural mesh, on the institution's hardware, under the institution's key.
Propagation — the loop closes
intra-AIRAWAT Meridian sync to other research nodes; optional pattern publication to NSM Phase 3 federated research mesh.
What just happened, in one paragraph
In ~120 milliseconds, AdiOS verified Prof. Saraswati Iyer's identity, ran 12 of 421 compliance controls, retrieved prior validated decisions from the institution's own memory, surfaced a recommendation in the language HPC professionals already use, took Prof. Saraswati's validation, and turned that validation into a permanent pattern. No customer data left the institution. By architecture.
See the HPC Journey live, in 30 minutes.
Book a 30-min demo of AdiOS for HPC. The 3-node cluster boot + HPC-specific BFSI-style end-to-end loop happens in real time.