When the ticket is created
Triage
Smart intake and classification. The ticket is checked against installed Application Packs, classified at module level and routed to a named resolver.
NIRA · Neev Intelligent Resolution Agent
Organizations can elevate their support capabilities without expanding headcount. NIRA automates the manual effort of ticket routing and historical research, empowering engineering teams to resolve issues faster. This translates to a significant reduction in required handling hours and a more agile, cost-effective operation.
< 7 sec
p50 classification latency
15–20
tickets per day, per engagement
4
independently swappable architecture tiers
Incident · P3Triage · on creation
Active mode · approval required
Illustrative interface
How NIRA works
NIRA acts at the three moments where intelligence changes the outcome of a ticket: when it arrives, while it is being worked, and when it closes.
When the ticket is created
Smart intake and classification. The ticket is checked against installed Application Packs, classified at module level and routed to a named resolver.
Any point during resolution
Analyst-triggered re-analysis. As the picture changes, the consultant can ask NIRA to re-evaluate the ticket and its resolution hints on demand.
On closure
Automated knowledge-article generation. A structured KB article is created from the resolved ticket and embedded, ready to inform the next similar ticket.
Between touchpoints: no ambient monitoring, no interruption.
Business impact
+54%
productivity per hour worked (0.84 → 1.29 weighted points/hour)
136.8 hrs
saved per month; 1,641.6 hours annualized
15–20%
modelled savings per year
Weighted points per hour worked
Ticket-handling hours per month
Platform flow
Six stages, from the moment a ticket is created in your ITSM tool to the knowledge article it leaves behind.
Stage 1
The ITSM connector normalizes the ticket to a canonical ticket model.
Stage 2
Is the ticket in scope for an installed Application Pack? Out-of-scope tickets pass through untouched.
Stage 3
Embedding, hybrid similarity search and an LLM call under Prompt Containment.
Stage 4
Classification plus the resolver registry produces a named person, never just a group.
Stage 5
The consultant agrees or corrects, and can refresh the analysis on demand.
Stage 6
On closure, a structured KB article is generated and embedded.
Every event is assessed independently. An append-only ACTION_LOG provides audit, not control flow — so the log records what happened without steering what happens next.
Architecture
Intelligence is built once. Application knowledge, ITSM plumbing and engagement settings each sit in their own tier, so any one can change without rebuilding the others.
Tier 1
Platform-agnostic intelligence, built once: classification, embeddings, retrieval and prompt logic.
Tier 2
Domain intelligence for each enterprise application.
Tier 3
Swappable plumbing to ITSM and integration platforms.
Tier 4
Per-engagement settings: scope mapping, resolver registry and field mappings.
Why it matters
| Capability | ITSM-native AI | NIRA |
|---|---|---|
| Application-specific classification | None | Deep, module-level (Oracle, EBS, Salesforce) |
| Resolution knowledge retrieval | Basic KB search | Vector similarity with curated resolutions |
| Cross-engagement intelligence | None (single tenant) | Compounding across the Neev portfolio |
| Multi-ITSM normalization | Own platform only | Platform-agnostic canonical model |
NIRA’s data-isolation rules apply throughout: every client deployment is fully isolated, and cross-client sharing is not permitted unless separately contracted, fully anonymized and reviewed by both clients.
Security architecture
NIRA’s security model rests on three pillars: data isolation and zero exfiltration; protection of IP, encryption and credentials; and operational safety controls on the LLM itself.
1. Data isolation
All tickets, embeddings, prompts and LLM responses stay inside the client boundary. Inference only — no client data is ever used to train a base LLM.
Every client deployment is fully isolated. Neev’s own core-engine IP is architected to contain zero client data.
If an external LLM or embedding endpoint is used, that transit point requires TLS 1.3 and a documented review of the provider’s data policy.
Not permitted unless separately contracted, fully anonymized and reviewed by both clients.
2. IP, encryption & credentials
Pack JSON, prompt templates and keyword dictionaries use native encryption (Oracle TDE / auto-encryption). Application-level encryption (DBMS_CRYPTO) protects AGENT_CONFIG.
The curation pipeline and scoring algorithms ship compiled (Cython/PyInstaller); function and variable names are obfuscated in deployed packages.
Widget platform shells deploy as managed apps (for example, Jira Forge) and are not decompilable by the client.
All credentials live in OCI Vault, AWS Secrets Manager or Azure Key Vault, referenced via the orchestration credential store. Never hardcoded; rotated per client policy.
Header-based API key for same-network calls; OAuth 2.0 Client Credentials for cross-network access.
Non-exclusive, revocable license with a non-reverse-engineering clause. IP obligations survive termination.
3. LLM safety controls
ACTION_LOG records every action, timestamped. CLASSIFICATION_LOG records who, what, when, why and how — including the CEIL patterns used.
Immediate halt on automated actions if misclassification risk emerges.
P1/P2 tickets are never auto-assigned without ≥95% confidence.
Governed Autonomy
NIRA moves through three operating modes. Authority widens only as accuracy is proven.
01
The agent classifies. Humans execute every action.
02
Recommends and assigns. Human approval required.
03
Executes directly, within confidence thresholds.
Stateless evaluation throughout: each event is judged independently, so no hidden state can quietly widen the agent’s authority.