Skip to content
Carlos Abreu
← Selected work

Agentic Private Credit / AI & trust

Agentic Private Credit: AI that investigates, verifies, and escalates

One contextual AI experience coordinating specialized capabilities over client-scoped sources, deterministic tools, evidence services, and human-reviewed records.

Product concept

Product direction only. No live agent or deployed capability is claimed.

The problem

Useful financial agents must verify evidence, respect entitlements, avoid unsupported legal conclusions, and escalate unresolved interpretation.

My responsibility

  • — Orchestrator and specialized capability requirements
  • — Deterministic tool-use hierarchy
  • — Extraction QA and evidence verification
  • — Entitlement and publication guardrails
  • — Layered evaluation design

What we’re connecting

  1. 1Private-credit user
  2. 2Contextual orchestrator
  3. 3Specialized capabilities
  4. 4Evidence service
  5. 5Entitlement gateway
  6. 6Client-scoped data and tools
  7. 7Human-reviewed records

Sanitized representative artifact

Decisions, alternatives, and trade-offs

07 / AI & trust

Private data boundaries belong in the architecture

Private agreements are client-owned; global AI search can cross boundaries.

Client-scoped sources and anonymized market aggregates.

11 / AI & trust

Professional judgment is part of the product model

A calculable score is not automatically a legitimate assessment.

Structured facts + deterministic factors + attorney judgment.

12 / AI & trust

Feedback loops require lineage by design

A reviewer correction is useless upstream if its origin is lost.

Track payload identity, source path, model version, original value, correction, evidence, and reviewer action.

16 / Workflow

A successful API call is not a successful experience

A success message appeared without a usable score or destination.

Treat missing score or URL as an incomplete outcome.

Extraction QA: inspect, verify, escalate

Sanitized representative artifact · Illustrative data

Extracted value

6.50x

Confidence: High · Illustrative

Validation / evidence

Pass

§6.12 · recreated source reference

  1. 1. Inspect extraction
  2. 2. Check required-field rules
  3. 3. Retrieve evidence
  4. 4. Compare with source
  5. 5. Check consistency
  6. 6. Classify issue
  7. 7. Recommend action
  8. 8. Escalate ambiguity

Guardrails before autonomy

  • No unsupported legal conclusions
  • No cross-client private-data leakage
  • No silent changes to published records
  • No automatic publication of judgment-sensitive outputs
  • Distinguish inference from calculation
  • Full evidence and tool trace
  • Requires legal review is a valid result

Evaluate the layers separately

  1. 01 / Rule correctness
  2. 02 / Evidence support
  3. 03 / Tool and workflow behavior
  4. 04 / Operational outcome
Evidence-supported claim precision
[Insert verified metric]
Escalation precision
[Insert verified metric]
Critical-error escape rate
[Insert verified metric]
Review-time reduction
[Insert verified metric]
User override rate
[Insert verified metric]

What I learned / what comes next

  • Deterministic functions first, structured facts second, retrieval when interpretation is needed.
  • Requires legal review is a valid result.

Verify rule correctness, evidence support, tool behavior, and operational outcomes before increasing autonomy.

Examples are recreated or sanitized to protect confidential company, client, and legal information. They preserve the product problem, decision logic, and working method without reproducing proprietary content.

Inspect product artifacts