Dhancept← BACK TO HOMEResponsible AI, by design
Dhancept's AI Swarm is engineered against the world's leading responsible-AI standards — the OECD AI Principles, the NIST AI Risk Management Framework, ISO/IEC 42001, and the EU AI Act — and, for the Indian market, in step with SEBI's disclosure and conduct norms for investment research — so that every signal, score, and recommendation our agents produce is trustworthy, explainable, and accountable by design, not by accident.
Seven Principles, Engineered In
These seven principles guide every model we ship — from a simple data summarizer to the AI agents that generate trade ideas.
Every AI agent in the Dhancept Swarm is built on the premise that trust must be earned through verifiable behavior, not assumed. Trust is treated as a foundational requirement, engineered into the system from day one rather than retrofitted after launch.
- We utilize a Model Registry to govern our primary AI systems, with a named owner and risk classification recorded for each
- Every report can be traced back to the specialist agents, data sources, and skills that produced it
- A model that fails or times out automatically opens a tracked incident and can trip a circuit breaker before it reaches more users
AI is designed to extend human decision-making, not substitute for it. Every analyst, trader, and researcher using Dhancept retains final authority — the system surfaces evidence and reasoning, but the human stays in the loop for any consequential action.
- All AI-generated trade ideas are clearly labeled and require explicit human confirmation
- Autonomous actions are scoped narrowly and bounded by pre-declared limits
- A one-click path to a human reviewer is available for every completed report
We believe the right response to AI risk is engineering discipline, not avoidance. New capabilities are shipped continuously, but every release passes through a structured risk-classification gate calibrated to the potential impact of the use case.
- Every AI model in production is catalogued with a Low, Medium, or High risk tier
- Low-risk internal tools ship fast; high-impact decision systems undergo deeper review
- Every model's lifecycle - registration, promotion, deprecation - is recorded as an auditable event
We hold ourselves to the standard that an AI-driven recommendation should never be shaped by who a user is rather than what the data shows. We are actively developing frameworks to track bias and fairness across our AI outputs, and are designing our architecture to support demographic parity tracking as adoption grows. Sensitive personal attributes are scrubbed from every prompt before it reaches a model.
- We are building an opt-in (off by default) pathway for users to voluntarily contribute self-declared demographic data - see Settings → Fairness Audit Participation - to help us evaluate disparate impact as our user base grows
- Sensitive personal attributes are automatically stripped from every prompt before it reaches a model
- Our roadmap calls for publishing disparate-impact statistics only once a demographic group reaches a meaningful, privacy-protective sample size
Accountability for an AI system's behavior rests with the team that builds and deploys it — never with 'the algorithm.' Each model in production has a named owner, a documented purpose, and a clear escalation path when something goes wrong.
- Every registered model has a named owning team and a board-level policy reference
- Structured incident reports capture root cause, remediation, and preventive action
- Grievances tied to an AI decision are logged, tracked through review, and marked resolved with a visible status trail
Black-box outputs are not acceptable for decisions that affect a user's capital or strategy. Every AI-driven recommendation on Dhancept ships with a plain-language explanation of the factors that drove it, so users can evaluate — and contest — the reasoning.
- Every report shows which specialist agents and skills contributed to the conclusion
- Findings are always accompanied by a human-readable rationale, not just a score
- Users can challenge or request a re-review of any AI-influenced outcome
AI systems are stress-tested before they are trusted. We design for graceful degradation: when a model's behavior drifts outside expected bounds, the system automatically falls back to safe, rule-based behavior and flags the event for human review.
- Automated drift detection monitors latency and failure rate on every model call in real time
- Circuit breakers suspend a model's output and reroute to a fallback provider when anomalies are detected
- Every failure is logged as an incident with root cause and remediation, reviewable by an administrator
How We Operationalize It
Principles only matter if they're enforced by infrastructure. Here's the compliance engine running behind every AI system on the platform.
A living registry of every AI system in production, each catalogued with an owning team, purpose, and risk classification for audit purposes.
- Central ledger of model lifecycle events (registration, promotion, deprecation)
- Every model recorded with a Low / Medium / High risk classification
- Internal governance policy reference recorded per model
Every model call's latency and outcome is recorded in real time, so a provider drifting outside normal bounds is caught automatically rather than discovered by a user.
- Real-time latency and failure-rate tracking across every AI model call
- Automatic circuit breakers that suspend a model and reroute to a fallback provider
- Defined automatic fallback to an alternate model provider during an outage
A structured, risk-calibrated internal review process covering input data, model logic, and output behavior. Independent, third-party audits and formal consent/provenance verification are part of our compliance roadmap as the platform scales, rather than active guardrails today.
- Three-dimensional reviews: data quality, model & algorithm, and output & behavior
- Internal self-assessments recorded per framework (OECD, NIST, EU AI Act, data privacy, demographic parity)
- Third-party audits and periodic re-certification are on our compliance roadmap, proportional to risk level
A formal, time-bound incident process ensures that any AI failure — bias, drift, or unintended behavior — is caught, reported, and remediated quickly.
- Severity-tiered reporting with defined response windows (hours, not weeks)
- Root-cause analysis and preventive-action tracking on every incident
- Aggregated, anonymized risk intelligence to catch systemic patterns early
Dhancept doesn't train or fine-tune its own models — every request runs through vetted third-party model providers, with personal data scrubbed before it reaches them or is stored.
- No proprietary model training occurs, so there is no training dataset to govern
- Every prompt is automatically scanned for common PII patterns - email addresses, phone numbers, card numbers, and SSN-format numbers - before it reaches a model provider; broader entity-level detection (e.g. names, addresses) is on our roadmap
- Data minimization — only what's necessary for the request is ever sent
Users always know when they're interacting with AI, always have a path to a human, and always retain control over consequential decisions.
- A clear AI-generated-content disclosure attached to every report, on-screen and in exports
- Accessible escalation to a human reviewer for any AI-influenced outcome
- Mandatory human-in-the-loop checkpoints for medium- and high-risk decisions
Trust isn't a feature. It's the foundation.
As AI becomes core to how capital is researched, allocated, and managed, we believe the platforms that win will be the ones users can verify, not just believe. Dhancept's responsible-AI program is reviewed and updated continuously as global standards evolve — because a framework that doesn't keep pace with the technology it governs isn't a safeguard at all.