How it works
Deterministic first. LLM only when it earns it.
Cognance is not a wrapper around a model call. It is a governed decision pipeline: exact matches resolve instantly and for free; genuinely ambiguous cases reach the reasoner; anything below the confidence threshold routes to a human — never auto-rejected.
The decision state machine
Every request flows through a small, auditable graph. Each node has a single responsibility, and the routing between them is explicit.
Deterministic check
Normalize both name strings and compare. On an exact match, the case is approved on the fast path with no model call — zero marginal cost, single-digit-millisecond latency.
LLM reasoner
Only ambiguous pairs reach the reasoner. A foundation model on Amazon Bedrock, run with deterministic settings, returns a structured decision with a confidence score and a plain-language reasoning log — for example, recognizing that 'Mohammed' and 'Mohd.' are common abbreviations.
Confidence routing
Above the confidence threshold, the case is auto-approved as AI-verified. Below it, the case is routed to human review. The agent never auto-rejects — in KYC, a wrong auto-approval is a compliance and fraud risk, so the system is deliberately cautious.
Human fallback + learning
A human resolves the held case. That decision is masked and appended to the golden dataset as a labeled example — so the next similar case can be auto-resolved. The loop compounds.
The pipeline
A serverless AWS architecture — the same event-driven pattern used to run regulated systems at scale.
Client LOS / CBS
Sends an application reference and two name strings
Managed API layer
Authenticates and ingests the request
Durable queue
Decouples ingestion from processing
Decision engine
Runs the governed decision state machine
Foundation model
LLM reasoning on Amazon Bedrock when needed
Audit store
Persists the immutable decision record
Webhook
Returns the decision to the client LOS
The golden dataset — a moat that compounds
Every human-review decision is masked at write time — the raw names are never stored — and appended to a proprietary golden dataset as a labeled example. Today it holds 46 labeled entries, seeded from real decided cases and growing with every review.
The more edge cases operators resolve, the better the agent gets at the Indian name variants that trip up exact-match tools — initials, abbreviations, reorderings, transliteration differences. Accuracy benchmarking is deliberately withheld until the dataset reaches a statistically meaningful size, so no accuracy figure is claimed before it can be earned.