Evals for Kalepa

Document classification

Correctly identifying document type across hundreds of formats within a single submission packet, including mixed and mislabeled attachments.

Field extraction fidelity

Pulling named insured, exposures, limits, and loss history values with correct units, dates, and entity attribution.

Loss run and SOV parsing

Handling the harder structured artifacts — multi-year loss runs and large schedules of locations — without dropping or merging rows.

Degraded input handling

Behavior on scans, partial packets, and unreadable pages: surfacing what could not be extracted instead of silently inventing values.

Conflict and duplicate detection

Recognizing the same risk arriving via different brokers or name/address variants and blocking auto-clearance.

Completeness verification

Determining whether a packet has what the workflow requires before it advances, and naming the specific gaps.

Sanctions screening

Screening named parties and reporting matches with enough identifying detail for a compliance reviewer to adjudicate.

Producer appointment and routing

Confirming the submitting producer is appointed and routing to streamlined or automated paths per the configured workflow.

Appetite and guideline matching

Aligning a submission to written appetite and underwriting guidelines, including in-appetite, out-of-appetite, and referral outcomes.

Bind-likelihood prioritization

Ordering the queue by likelihood to bind and stating what drove the ranking.

Declination and referral signaling

Producing an actionable disposition rather than a bare score, with the guideline clause that triggered it.

Cross-line consistency

Holding triage behavior stable across the commercial lines Kalepa claims to cover.

Flag assignment and severity

Assigning red, yellow, or preferred consistent with the underlying finding rather than over- or under-escalating.

Evidence attribution

Tying each flag to the document, field, or external source it came from so an underwriter can verify it.

Exposure and controls analysis

Reading protection features, occupancy, and operations into risk-relevant conclusions.

Consolidated view assembly

Bringing terms, communications, and analysis into one coherent view without contradicting itself across panels.

Portfolio aggregation

Rolling submission-level exposures and dispositions into accurate book-level totals and concentrations.

Steering signals

Surfacing where the book is drifting relative to appetite and targeted mix.

Performance attribution

Connecting portfolio outcomes back to the selection decisions that produced them.

Rule capture and adherence

Operating against an insurer's own underwriting logic and behaving predictably when that logic is ambiguous or unstated.

Auditability of outputs

Leaving a reviewable trail of what was concluded, from which inputs, at what time.

Human handoff and escalation

Handing judgment calls to an underwriter with the context needed to decide, rather than resolving them unilaterally.

Embedding in existing process

Fitting into the incumbent workflow and system of record without requiring the underwriter to work outside it.