Decision Integrity Layer for Financial Institutions
Shadow-mode AI infrastructure for detecting compromised financial decisions before irreversible loss occurs.
Fraud systems detect malicious actors. Sesim detects when a real customer may no longer be making a free decision.
TRY 180,000 transfer to new beneficiary
The actor can be legitimate while the decision is compromised.
In coercion-style transfer risk, the device, login, and customer identity can all look clean. Sesim adds a decision integrity layer that observes pressure, urgency, stress movement, and behavioral drift beside the existing fraud stack.
Clean fraud profile
Known device, valid login, real customer, authorized transfer.
Compromised decision
Active call pressure, unusual urgency, new beneficiary, rising hesitation.
Operational recommendation
Soft pause, safer-channel reconfirmation, or silent escalation to fraud care.
Decision Integrity is layer 1. The full stack is a four-layer voice-first platform.
Our patent application covers four modular layers. Each can be adopted independently, sold separately, and runs on a shared voice integrity engine (anti-spoofing, ZKVP, NIST PQC, edge inference) and audit trail. Decision Integrity (layer 1) is the wedge available today.
Layer 1 — Decision Integrity
Catch the moment a real customer is no longer making a free decision.
- Primary buyer
- Anti-fraud / financial-crime / customer-protection (banks, insurance, telco call centres)
- Patent moat
- Coercion / duress detection + silent alarm (claims 20, 47); behavioural biometry & prosody (claims 19, 46, 60)
- Stage
- Live — demo and deterministic engine
Layer 2 — Cryptographic Voice Identity
Cryptographically anchored voice identity — live challenge, anti-replay, anti-deepfake by patent design.
Passwords leak. Devices get cloned. Deepfakes imitate. A live cryptographic voice challenge cannot simply be replayed.
- Primary buyer
- Mobile banking, ATM, call-centre IVR, telco caller-auth, e-government, insurance phone-claims
- Patent moat
- Voice-biometric system + NFC + ZKVP/zk-SNARKs + NIST PQC (claims 1–15, 21–25, 41, 51, 43, 50, 58)
- Stage
- In product / pilot-ready
Layer 3 — Cross-Layer Voice Transaction
Three signals from one voice intent — who, under what pressure, what action — bound to a tamper-evident distributed audit chain.
Voice payments are easy. Voice payments under coercion or impersonation are hard. A single utterance becomes evidence — three cross-layer signals plus a 3-validator Proof-of-Voice consensus on an IPFS/DHT-anchored audit chain.
- Primary buyer
- Bank mobile / POS, ATM, healthcare consent, public services, transport turnstiles, education / exam integrity, e-commerce, call-centre auth, hospitality (9 sectoral claims)
- Patent moat
- NLP + voice smart-contract + sectoral usage (claims 26–36, 39, 40, 52); 3-factor cross-layer single-utterance handshake (41, 42); 3-validator Proof-of-Voice consensus (50); IPFS/DHT off-chain audit (56); GDPR/KVKK lifecycle (38)
- Stage
- Pilot-ready — partner-led
Layer 4 — Bias-Fair Inclusive Access
Multi-modality access with continuous fairness audit per cohort — every cohort counts, every cohort tracked.
Screen readers help one user. Bias hides in everyone else. Sesim Layer 4 monitors fairness deltas across accent, age, gender and language-switch cohorts — and pauses production for any cohort that breaks the threshold.
- Primary buyer
- Public sector (e-government, municipalities), banks (compliance), insurance, education, NGOs · covers 4 disability segments (visual, hearing, motor, cognitive) + elderly cohort
- Patent moat
- 4-modality accessibility incl. cognitive (claim 16); voice + haptic feedback (25); visually-impaired ATM/kiosk integration (27); multi-language code-switch (49); bias detection + fairness mitigation across demographic cohorts (53); inclusive cross-layer (109)
- Stage
- Pilot-ready — partner-led
A typical bank customer starts with Layer 1, then expands into Layer 2 (call-centre auth) and Layer 3 (voice-confirmed payments). Layer 4 can be co-pursued with public-sector or compliance teams. Net dollar retention model: 3–4× ARR over three years.
Synthetic operating console for a coercion-transfer review.
A bank risk team sees a clean fraud engine result, but Sesim flags compromised decision context and produces reason codes, timeline evidence, analyst notes, and audit logs.
TRY 180,000 transfer to new beneficiary
Fraud profile clean; known mobile device and normal login.
Customer remains on an active call during beneficiary setup.
Hesitation and contradiction increase during confirmation.
Soft pause before irreversible transfer execution.
Sesim does not replace the fraud engine. It covers the decision blind spot.
Optimized for malicious actor, credential, device, location, and session anomalies.
Optimized for pressure, autonomy, urgency, stress trend, and contextual drift.
Every recommendation is backed by a reviewable signal path.
Pressure Indicators
Authority cues, urgency language, active call context, repeated prompting.
Behavioral Deviation
New beneficiary, compressed confirmation time, unusual transfer path.
Emotional Stress
Rising voice stress trend and hesitation during sensitive prompts.
Recommendation
Soft pause, safer-channel reconfirmation, or silent care escalation.
Protect vulnerable customers without exposing them to the person applying pressure.
The first wedge is not generic fraud automation. It is a care-aware review layer for moments when an elderly or vulnerable customer may be guided through a legitimate-looking transfer under fear, urgency, or social pressure.
Active call pressure
The customer stays on a call while adding a new beneficiary and confirming a high-value transfer.
Fear of contradicting the caller
The customer appears compliant but hesitates when asked to explain the transfer purpose.
Safe reconfirmation
The bank introduces a pause and asks for confirmation through a safer channel without warning the coercer.
Audio-first guidance and confirmation flows for transfer review on the customer side.
Alternative haptic / vibration confirmation cues and visual prompts replace audio-only feedback.
Voice-only authentication and confirmation paths so a transfer review never requires precise touch input.
Simplified-command paths and slower-paced confirmation so customers under cognitive load are not pushed through dense decision flows.
Seven decision signals feed one explainable Decision Integrity Score.
Externally it is one explainable score with reason codes and an intervention ladder. Internally it is a modular, independently updatable signal engine. The live demo shows a representative subset; the rest activate on the roadmap with institution data.
Decision Ownership (DOS)
LiveIs this decision the customer’s own, or externalized under pressure?
Intent
LiveDoes the stated purpose match behavior and transaction history?
Velocity
LiveAbnormal decision speed versus the customer’s own baseline.
Context
LiveAuthority pressure, urgency, active-call, and channel context.
Audit
LiveImmutable reason-code and decision trace behind every score.
Deepfake
RoadmapSynthetic voice or video manipulation at the decision moment.
Regret
RoadmapPost-decision regret signal used to tune false positives.
The category itself — detecting elevated stress and coercion-style speech, triggering an automatic protection protocol and a silent alarm path for users under threat — is part of our patent-pending IP. The behavioural-biometry primitives (prosody: pitch, stress, rhythm, pace; pause patterns, word choice, emotional tone) that feed our decision signals are covered by the same application.
A transparent, deterministic score — not a black box.
Sesim does not return an opaque number. Each decision signal is normalised to a 0–1 degradation, combined with published weights, and subtracted from a perfect score. The same engine powers the live API and this demo.
1 · Signals in
Transaction, context, and behavioural signals (active-call pressure, confirmation latency, beneficiary anomaly) arrive via a read-only event.
2 · Per-signal scoring
Each live signal is scored independently and returns its own weighted contribution — fully inspectable.
3 · Decision Integrity Score
Contributions are fused into one 0–100 score with reason codes and an intervention tier.
One decision layer, several compromised-decision patterns.
The wedge is coercion and social engineering, where the actor looks clean but the decision is not. The same signal engine generalises across related patterns — pilot scope is defined narrowly with your team.
Authorised Push Payment / coercion
A pressured or elderly customer is guided through a legitimate-looking transfer while on an active call.
Boiler-room investment fraud
No single transfer looks wrong; decision integrity erodes across weeks of repeated “investment” payments.
Account-takeover variants
New-device login, session/behaviour mismatch, social-engineered step-up — here the fraud engine leads and Sesim adds decision context.
Deepfake-assisted manipulation
Synthetic voice or video used to apply authority pressure at the decision moment (roadmap signal).
Deploy beside existing fraud workflows before touching production decisions.
Sesim starts in shadow mode: observe high-risk journeys, generate explainable decision-integrity signals, compare against existing outcomes, and only then define production-safe intervention policy.
Receives device, login, session, and transaction risk outcomes.
Observes selected journeys without blocking customer flow.
Generates DIS, AUPA, YUDA, pressure indicators, and reason codes.
Stores recommendation, reason code, analyst note, and review status.
Routes sensitive cases to fraud, compliance, or customer care teams.
Built for the bank reflex: your data does not leave your environment.
Sesim is designed to run inside your perimeter, beside the fraud stack — no data egress, no core-banking change. The points below are the target pilot posture; exact commitments are fixed in writing with your security and architecture teams at kickoff.
In-VPC / on-prem inference
Model inference runs inside your environment. Customer data does not leave the bank and is not written to our logs.
Edge-ready voice analysis
Voice signals can be analysed on-device or at the edge — no raw audio leaves the customer endpoint when the deployment requires it.
Read-only event stream
We consume a copy of payment / fraud events via API or event bus. No write path to core banking and no schema change.
No customer-flow latency
During the pilot scoring runs in parallel / async, so it adds no latency to the customer transaction path.
Zero-knowledge voice verification (ZKVP, zk-SNARKs)
Voice is verified via zk-SNARK proofs with client-side proof generation — the raw audio is never revealed, preserving GDPR / KVKK alignment by design (patent-pending IP).
Anti-spoofing & liveness
Liveness detection plus deep-learning classifiers, augmented with spectral analysis, reject replay and synthetic-voice (deepfake) inputs (patent-pending IP, integrated during pilot).
Post-quantum cryptography (NIST PQC)
NIST-standard primitives — CRYSTALS-Kyber, CRYSTALS-Dilithium, FALCON, SPHINCS+ — protect verification traffic against future quantum-class threats.
Replay & brute-force resistance
Dynamic challenge–response with a 10,000+ random challenge pool, time-stamp protection against replay, and exponential-backoff rate limiting against credential stuffing (patent-pending IP).
GDPR / KVKK-aligned data lifecycle
AES-256 off-chain storage with SHA-256 timestamped pointers anchored to the audit trail; first-class data-erasure and key-destruction paths so right-to-be-forgotten requests do not need to mutate the chain.
Immutable audit trail
Every score carries a reason-code and decision trace — answerable for internal audit and regulators; optional blockchain-anchored audit (with IPFS / DHT off-chain payloads) available.
Synthetic demo, real pilot
This website and dashboard use synthetic data only. A real pilot operates on your data, in your environment.
Narrow, defined scope
A pilot covers a deliberately narrow decision surface (e.g. outbound transfers) — never “all fraud”.
Start with zero customer impact. Prove value before touching a single decision.
Sesim runs entirely beside your stack. In stealth / report-only mode it produces recommendations and audit logs but takes no action — so a pilot carries no operational or customer risk.
Shadow Mode
Observes selected high-risk journeys in parallel with the fraud engine. No customer-facing change.
Stealth / Report-Only
Generates recommendations and logs only — never intervenes. Neither the customer nor a coercer can tell it is running.
Active Policy (post-pilot)
After review, selected intervention tiers are switched on with your fraud, compliance, and care teams.
Run Sesim in report-only mode for 30 days. At the end you receive a "what Sesim would have stopped" report — flagged compromised-decision transfers, estimated prevented loss, and false-positive review — with no action taken on any live transaction.
A single scoring call. Your interface, your fraud stack, your intervention flow.
Sesim is a headless decision-integrity engine. Send transaction and context signals to one endpoint and receive a Decision Integrity Score, reason codes, and a recommended intervention tier. Keep your own dashboard and workflow — integration moves from weeks to days.
POST /api/score
{
"transaction": { "amount": 180000, "currency": "TRY", "beneficiary": "new" },
"context": { "active_call": true, "channel": "mobile", "age_band": "70+" },
"signals": { "confirmation_latency_ms": 1400, "hesitation_events": 3 }
}200 OK
{
"dis": 38,
"tier": "soft_pause_care_call",
"reason_codes": ["authority_pressure", "stress_rising", "behavioral_anomaly"],
"recommended_action": "Soft Pause + Customer-Care Callback",
"mode": "report_only"
}Live & callable
This is a real, deterministic engine — GET /api/score for the contract, POST to score.
Headless
No UI lock-in. Keep your existing fraud console and case management.
Explainable by default
Every score returns a weighted signal breakdown, reason codes, and a tier.
A clear path from today’s shadow-mode demo to a category-defining decision-integrity network.
Everything live today runs without touching production decisions. Pilot-phase and vision capabilities activate with institution data and review.
Discovery with fraud, compliance, innovation, and customer protection teams.
Shadow-mode simulation on selected synthetic or anonymized transaction journeys.
Decision integrity lift, false-positive review, and intervention policy proposal.
Review a synthetic coercion-transfer scenario with your risk team.
We are looking for 3 shadow-mode pilot conversations with banking and fintech teams.
Request Shadow-Mode Pilot
Tell us who should evaluate Sesim. We will reply with the shadow-mode pilot brief, synthetic scenario, and dashboard walkthrough.
Prefer email? Contact admin@sesim.net