Every product is the same platform, recombined.
These are the components we have already built. Pick a solution and step through exactly how it is assembled — the problem it answers, the specialized agents it draws on, which of your systems it connects to, and the workflow wrapped around it. Recomposition, not a rebuild.
The methodHow a product gets built — in six moves
What the six moves draw on
Integration is repeatable because it is organised by system class, not by project. Name your system — the connector class is already certified.
Core banking & transaction systems
Data platforms & warehouses
Communication channels
Document & content repositories
Workflow, ticketing & case management
CRM & customer engagement
Identity, access & entitlement
Risk, compliance & regulatory
External & third-party data
Automation & RPA
AI runtime & knowledge stores
Observability & audit sinks
Seven modules in four layers. M1 maps one-to-one onto the twelve system categories; M2 is the core; M3–M5 are the auxiliaries that serve it; M6 and M7 are cross-cutting architecture — the guardrails, human gates and audit that make every output defensible.
Connector fabric
The integration layer — and it maps one-to-one onto the twelve client system categories. Every category has a certified connector class: auth, schema mapping, rate limits, retries and lineage solved once, reused by every agent.
- One certified connector class per system category — not per project
- Credential vaulting with per-connector scopes
- Schema-drift detection with lineage capture
- Read/write separation — every write passes a human gate
Specialised agent selection
The core of the platform: route each unit of work to the right specialized agent — and, beneath it, the right model for that task. This is where token efficiency, long-context handling and multi-agent orchestration are decided.
- Task-level routing to the agent whose framework fits the job
- Model routing beneath the agent: small models for extraction, larger only where reasoning earns it
- Token efficiency — retrieval and summarisation collapse context before it reaches the model, so cost scales with work, not with data volume
- Long context handled by decomposition: state objects and scoped retrieval instead of one enormous prompt
- Multi-agent orchestration and swarming — many specialized agents run in parallel over partitioned work, then reconcile
- Open-weight default, hosted burst optional; published accuracy per task
Knowledge-base selection
Auxiliary — supplies the right corpus for the question at hand: policies, circulars, schemas, product manuals, indexed with retrieval tuned per agent.
- Per-domain corpora with access scoping
- Chunking and embedding tuned per document class
- Freshness policies and re-index triggers
- Source citation on every grounded answer
Specialized prompt libraries
Auxiliary — versioned, domain-tuned prompt assets: the encoded output of the ~100 iterations an enterprise would otherwise run itself.
- Versioned and diff-able, tied to eval results
- Jurisdiction and language variants
- Role and persona framing
- Regression-tested before promotion
Memory & state management
Auxiliary — gives every unit of work a durable state object with an owned lifecycle, so agents can hand off, resume and swarm without losing the thread.
- Durable state per case, not per chat session
- Replay and resume after failure
- Short- and long-term memory scoping per agent
- API-driven, inspectable at every transition
Guardrails
Cross-cutting architecture — applied before an agent sees data and before any action leaves the platform, identically in every product.
- PII detection, masking and redaction on ingress
- Prompt-injection and jailbreak screening
- Allow-listed action catalogue per agent
- Confidence thresholds that force human review
- Mandatory human gates on consequential actions
Common observability
Cross-cutting architecture — one way to trace, evaluate, audit and report every agent in every product. The layer a bank's risk function actually buys.
- Step-level traces: prompt, retrieval, tool, decision
- Ground-truth evals and accuracy dashboards
- Immutable audit of every human approval
- Export to the bank's SIEM and warehouse
Generic capability is a commodity. Each agent carries its own framework — the framework is the IP; the agent is its packaging.
Document intelligence
Parse, classify and extract structured fields from enterprise documents at production grade.
Specialized because: Document-class-specific extraction schemas; confidence per field; layout + language robustness.
Legal clause extraction & mapping
Pull clauses from legal and regulatory documents; map mandatory vs non-mandatory obligations.
Specialized because: Obligation taxonomy tuned per jurisdiction — the capability an enterprise needs ~100 iterations to reach.
Document comparator
Compare versions, policies and contracts; flag material differences.
Specialized because: Materiality scoring, not diffing — what changed that matters.
Image intelligence
Compliance-grade image checks including letterhead, signature and Shariah screening.
Specialized because: Domain classifiers over generic vision; audit-ready verdicts.
Email listening & screening
Ingest, authenticate and triage inbound mail; screen scam and unauthorized senders.
Specialized because: Sender authenticity scoring; routing before reading costs money.
Email & response drafting
Draft jurisdiction-correct replies and notices for human approval.
Specialized because: Templated + generated hybrid; never dispatches without a gate.
Schema & DB intelligence
Connect any database; synthesize and understand tables, fields and relationships.
Specialized because: Metadata harvesting + semantic layer built automatically.
Text-to-SQL
Natural language to governed queries over live schemas.
Specialized because: Hybrid RAG + text-to-SQL framework: deterministic plan checker, self-correcting validator, read-only by construction.
ETL / transform agents
Describe, profile and transform tables into one standardized model with lineage.
Specialized because: Rule + pattern hybrid; lineage emitted as a first-class artifact.
Synthetic data generation
Generate governed, zero-PII test data preserving referential integrity and business rules.
Specialized because: Schema-driven generation with a five-state inventory lifecycle.
Anomaly detection
Rule-based and pattern-based detection across the full transaction population.
Specialized because: Full population, not 2–5% sampling; known and emerging deviations.
Entity enricher
Research and enrich merchants, offers and product entities from external sources.
Specialized because: Governed crawling plus compliance screening built in.
TasteGraph recommender
Transactions → affinity scores → ranked, explainable recommendations.
Specialized because: Two-tower embeddings; 2,888M customer-merchant affinities; sub-1.5s at scale.
Inference & decisioning
Score, rank and decide with explainability attached to every output.
Specialized because: Deterministic rules where required, models where they earn it.
Action & dispatch
Execute the approved action into the system of record — API or RPA.
Specialized because: Idempotent, gated, fully logged; RPA fallback for API-less systems.
Capability finds purpose only when mapped to a known problem
- A generic capability finds purpose only when mapped to a real, industry-known problem. “Document intelligence” sells nothing; “verify a law-enforcement freeze order in 30 seconds across 22 languages” wins a bank.
- The 100-iteration argument. LEANM distinguishes mandatory from non-mandatory legal clauses out of the box. An enterprise building this on a generic platform needs ~100 prompt iterations, domain experts and months of eval work to reach the same point — per use case, per jurisdiction.
- We take open source and make it fit for purpose. We do not rebuild wheels; we reuse what exists and innovate on top — architecture, evals, domain tuning. Said out loud: a strength, not an admission. It is why our cost base and speed beat both in-house builds and closed platforms.
Specialization is an economics argument too
Token efficiency
A specialized agent knows what it needs. Scoped retrieval and staged summarisation collapse context before it reaches the model, and small models handle extraction while large ones are reserved for genuine reasoning — so cost scales with the work done, not with the size of the data estate.
Long context, handled by decomposition
Rather than pushing an enormous prompt at a bigger context window, work is decomposed into durable state objects with scoped retrieval per step. Accuracy holds as documents and case histories grow, and nothing depends on a model vendor's context limit.
Multi-agent orchestration & swarming
Because agents are discrete and state is externalised, many can run in parallel over partitioned work and reconcile at a gate — full-population control checks, 50,000 notices a month, 421M transactions. Throughput scales horizontally; the audit trail stays single and complete.
Same modules, same agents — different workflow, different systems. Modules M1–M7 and the human-gate + audit layer apply to every one.
LEANM · Notice Management
Reads, verifies and routes every regulatory & law-enforcement notice — end to end.
Specialized agents
Client systems touched
Custom domain workflow
- Seven-stage case pipeline (F1 acquire → F7 dispatch)
- Mandatory DB-enforced human gate at F6 — nothing reaches a regulator without it
- Deadline / SLA clocks per regulator and jurisdiction
- Event-driven agent handoffs, each an audit record
Controls & Anomaly Detection
Continuous, full-population control monitoring with automated anomaly detection.
Specialized agents
Client systems touched
Custom domain workflow
- Nine-stage control lifecycle — design, operate, detect, resolve, assure
- Exception case management with multi-level approval and evidenced closure
- Standardized data model across source systems with quality checks
- Live control-effectiveness dashboards for audit
CXO Concierge
Ask the banking business anything — sourced answers, no analyst, no SQL.
Specialized agents
Client systems touched
Custom domain workflow
- 12-step question flow with retry loops and validator routing
- Persona-aware framing — tone, KPIs and suggested questions per role
- Read-only by construction: no INSERT/UPDATE/DELETE ever generated
- Row-capped execution with source citation on every answer
Test Data Hub
AI-generated, governed synthetic test data on demand — zero PII.
Specialized agents
Client systems touched
Custom domain workflow
- Five-state inventory lifecycle: generated → allocated → reserved → released → consumed
- Three-tier RBAC (org → product → project) with row-level security
- RPA-driven orchestration into core systems with retry/failure handling
- Audit-grade logging of every allocation and access event
Entity Enrichment (EMP)
Research, operate and distribute banking product entities at scale.
Specialized agents
Client systems touched
Custom domain workflow
- 13 discrete agents composed by one orchestrator across research → operations → distribution
- Compliance module per jurisdiction (RBI, IRDAI, MAS, Shariah)
- Entity lifecycle states with immutable audit records
- Unified API gateway serving structured entity data downstream
AI Marketplace (Personalised CX)
Convert transaction data into real-time, ranked, explainable decisioning.
Specialized agents
Client systems touched
Custom domain workflow
- Four lifecycle stages: cold start → active → intent → lapsed
- Six decisioning agents behind one Dynamic Lists API
- Fatigue monitoring and rotation with anti-repetition rules
- A/B experimentation built into the serving path
Read down a column to see how a product is assembled; read across a row to see how often a capability is reused.
| Specialized agent | LEANM | Controls & Anomaly Detection | CXO Concierge | Test Data Hub | Entity Enrichment | AI Marketplace |
|---|---|---|---|---|---|---|
| A1Document intelligence | ● | · | · | · | ● | · |
| A2Legal clause extraction & mapping | ● | · | · | · | · | · |
| A3Document comparator | · | · | · | · | · | · |
| A4Image intelligence | ● | · | · | · | ● | · |
| A5Email listening & screening | ● | · | · | · | · | · |
| A6Email & response drafting | ● | · | · | · | · | · |
| A7Schema & DB intelligence | · | ● | ● | ● | · | · |
| A8Text-to-SQL | · | · | ● | · | · | · |
| A9ETL / transform agents | · | ● | · | ● | · | · |
| A10Synthetic data generation | · | · | · | ● | · | · |
| A11Anomaly detection | · | ● | · | · | · | · |
| A12Entity enricher | · | · | · | · | ● | ● |
| A13TasteGraph recommender | · | · | · | · | · | ● |
| A14Inference & decisioning | · | ● | ● | · | ● | ● |
| A15Action & dispatch | ● | ● | · | ● | ● | ● |
The reuse is the argument: every engagement consumes proven capability and returns hardened capability to the registry.
Agents find purpose only when mapped to a known problem
Generic capability sells nothing
“Document intelligence” sells nothing; “verify a law-enforcement freeze order in 30 seconds across 22 languages” wins a bank.
The 100-iteration argument
LEANM distinguishes mandatory vs non-mandatory legal clauses out of the box. An enterprise building this on a generic platform needs ~100 prompt iterations, domain experts and months of eval work — per use case.
Open source, made fit for purpose
We reuse what exists and innovate on top — architecture, evals, domain tuning. Said out loud: a strength, not an admission.
Evidence, not adjectives
Framework-free by design
Frameworks add abstraction and micromanage state you don't need — and strip the control regulated workflows demand. We own every transition, API-driven. (We moved off LangChain/LangGraph on Controls and LEANM.)
Self-healing, storage-aware retrieval
How data is stored defines how it is retrieved; retrieval architecture, metadata design and self-correction are engineered together, not prompted together.
Event-driven agent handoffs
Every handoff is an auditable event; every notice a state object through a seven-stage pipeline with one mandatory, DB-enforced human gate.
One unique framework per agent
Example — the hybrid RAG + text-to-SQL framework (the Text-to-SQL agent, as used in CXO Concierge): schema-mapping agents, deterministic zero-LLM plan checker, self-correcting validator loop. 14+ agents behind one interface.
Deliberately not a drag-and-drop builder
Real workflows carry semantics
Ticketing, statement management, RBAC, multi-level approvals, SLA clocks and evidence trails — a no-code canvas cannot express any of it.
Customized per product
The same base agents, wired into each solution's regulated workflow. That is the domain build — and the reason our products survive an audit.
The same components, wired into your workflow
Every capability above already runs in production. For a given problem we recompose the right agents, connect them to your systems, and wrap them in your regulated workflow — assembled and hardened, not built from scratch. The connector work to your systems is the only genuinely client-specific effort.