Radha Kant Jha
Sets the firm's direction and stays close to delivery. Co-founded EduJha, an AI-assisted exam platform now running in production across multiple exam tracks.
AI-first doesn’t mean
AI-only.
A model is the smallest part of an enterprise AI system. What surrounds it decides whether the thing survives contact with production.
Most failed AI programmes are not model failures. They are missing data contracts, absent evaluation, and unclear accountability.
You are buying judgement.
Engagements are led by the people who designed and shipped the work below — not staffed out after the pitch. Small teams, senior by default.
Nisha Kumari
Leads architecture and the AI practice. Designed EduJha's model layer, which routes across several providers behind one interface so no single vendor sets cost or latency.
Devendra Jha
Owns the evaluation side of AI delivery — building the datasets, rubrics, and review processes that decide whether a model's output is fit to ship.
Tarun Kumar
Shapes what gets built and in what order. Focused on the interfaces where people exercise judgement over what an AI system proposes.
Software engineering
Models are one component. The system around them is still software, and it is held to software standards.
Data foundations
Lineage, contracts, and semantics — the difference between an AI demo and an AI system.
Security
Threat modelling for prompt injection, data exfiltration, and tool misuse from the first design review.
Governance
Model registries, approval gates, and decision records that survive an audit.
Observability
Traces across every reasoning step, tool call, and retrieval — not just request latency.
Evaluation
Offline and online evaluation as continuous infrastructure, not a launch checkpoint.
Human oversight
Explicit escalation paths, reversibility, and accountable review where the stakes justify it.
Responsible AI
Bias testing, disclosure, and data-handling boundaries defined before deployment.
Built for enterprise reality.
Pilots run on goodwill. Production runs on boundaries, evidence, and an operating model — which is where most enterprise AI work actually stalls.
Security
Zero-trust boundaries, secrets management, and adversarial testing of AI-specific attack surface.
Governance
Policy as code, model and dataset registries, and traceable approval workflows.
Scalability
Architectures that hold their shape from pilot cohort to enterprise-wide rollout.
Reliability
SLOs, graceful degradation, and deterministic fallbacks when a model is unavailable.
Compliance
Engineering that maps to your regulatory obligations and produces the evidence auditors ask for.
Observability
End-to-end tracing across services, data, prompts, and decisions.
Data privacy
Residency, minimisation, retention, and redaction enforced in the architecture, not the policy document.
Orvient does not list certifications it does not hold. Certification and attestation status is shared directly during procurement.
Talk to the people who would do the work.
No pre-sales layer. The first conversation is with the engineers who would run the engagement.