Nkosi Felix, Senior Data Platform Engineer
Also: what Nkosi Felix can do.
Each capability below is backed by a public page, lab, source file, test, evidence entry, or bounded system behavior. Unsupported claims are omitted.
Demonstrated by: NBCU/Universal Creative document-ingestion harness, Microsoft Graph/SharePoint delta-sync, validation routing.
Inspect:
Boundary: Does not expose employer-confidential data or claim unsupported production scope.
Demonstrated by: idempotent Kafka-to-PostgreSQL consolidation service with replay handling and DTO/schema validation at 4C Strategies.
Inspect:
Boundary: Public reference implementation uses synthetic data, not the original employer's proprietary pipeline.
Demonstrated by: ACCEPT/QUARANTINE/SYSTEM-failure routing (NBCU), DTO validation (4C), Field Change Order Evidence Tracker lab.
Inspect:
Boundary: Labs are synthetic; no real employer or customer data is used.
Demonstrated by: Grafana/Prometheus/Loki dashboards at 4C, documented ~30% reduction in debugging cycles; this site's own fail-closed chat/signal degradation.
Inspect:
Boundary: The ~30% figure is the only allowed metric for that role; no other metrics are implied.
Demonstrated by: contract-test gates at 4C (JUnit 5, Mockito, Spring Cloud Stream test binder); Acceptance Harness (7/7 tests); Air-Gap Messaging Ingest (26/26 tests).
Inspect:
Boundary: Test counts are from this project's own CI output, not a third-party audit.
Demonstrated by: this site's own chat — evidence ledger, generated manifest, policy engine, and adjudicator.
Inspect:
Boundary: Does not claim unrestricted chatbot intelligence or private data access.
Demonstrated by: a canonical chat router that answers identity/employer/feature questions with zero LLM calls.
Inspect:
Boundary: Covers a bounded set of public questions; unmatched questions fall through to a provider-assisted or fail-closed path.
Demonstrated by: live, provider-assisted answer paths used only when configured and healthy, gated by evidence/safety/confidence checks before anything reaches a client.
Inspect:
Boundary: Provider/model identity is never exposed in the public response; used only when selectable and healthy.
Demonstrated by: fail-closed adjudication, demonstrated by the unsupported-claim rejection example on the homepage.
Inspect:
Boundary: Rejection is deterministic and policy-driven, not a guess.
Demonstrated by: deterministic (no-LLM) resume-match and JD-fit-score logic reachable through the evidence-backed chat.
Inspect:
Boundary: Does not rewrite the resume, does not guarantee interviews or offers, does not store pasted job descriptions by default.
All labs are synthetic/public-safe. No real patient, grid, construction, customer, payment, employer, or private tracker data.
Demonstrated by: Patient Signal Intake Lab — schema validation, quarantine routing, review packets.
Inspect:
Boundary: Synthetic-only. No PHI. Not a HIPAA-compliance claim.
Demonstrated by: Grid Event Quality Monitor — timestamp validation, drift detection, degraded-telemetry handling.
Inspect:
Boundary: Synthetic project data only.
Demonstrated by: Field Change Order Evidence Tracker — document intake, duplicate detection, owner-approval boundary.
Inspect:
Boundary: Synthetic project data only.
Demonstrated by: Evidence-Grounded RAG Evaluator — unsupported-claim rejection, evidence gates, confidence/metadata, source-backed answers.
Inspect:
Boundary: Mirrors this site's own chat; not a demo of a separate production system.
Demonstrated by: Acceptance Harness — pure functions, no I/O, typed outcomes, 7/7 JUnit tests.
Inspect:
Boundary: Public-safe synthetic/reference implementation.
Demonstrated by: Air-Gap Messaging Ingest & Observability — Kafka-to-Postgres, duplicate handling, Grafana/Prometheus/Loki, 26/26 tests.
Inspect:
Boundary: Synthetic, public-safe data only.
Demonstrated by: ShopLabs Lakehouse Lab — Python/DuckDB, synthetic Faker-generated order data, Parquet export.
Inspect:
Boundary: No real customer, payment, or store data.
Demonstrated by: CareerAssistAI Scoring Pipeline — FastAPI service, tokenized keyword matching, transparent rationale.
Inspect:
Boundary: Synthetic inputs only; no private tracker or recruiter data.
Demonstrated by: resume-match and JD Fit chat capabilities — requirement-to-evidence mapping, recruiter-safe reply drafts.
Inspect:
Boundary: No invented claims or fake metrics; does not guarantee interviews or offers.
Demonstrated by: a dedicated Recruiter Packet (work authorization, availability, submission facts) and a separate Hiring Manager Proof page (architecture, tests, evidence, operations).
Inspect:
Boundary: No active clearance claimed; no degree-completion claim.
Demonstrated by: JD Fit Score's hard-blocker detection and gaps/verify-items output.
Inspect:
Boundary: Deterministic, no LLM call; does not mutate any real job tracker.
Demonstrated by: Location-Aware Weather Signal and Global Currency Basket.
Inspect:
Boundary: Not investment advice, not booking authority, not an official airline/provider partnership.
Demonstrated by: AI News Ticker — SpaceX / xAI-Grok / Anthropic-Claude / OpenAI-ChatGPT-Codex lanes, in-memory + seed cache fallback.
Inspect:
Boundary: Never backfills a missing lane with another lane's story; omits quietly instead.
Demonstrated by: a dictionary-swap language toggle with localStorage persistence, applied across every public page.
Inspect:
Boundary: Does not claim native Spanish fluency or a professional translation certification.
Demonstrated by: ai-index.json, llms.txt, sitemap.xml, claims-register.json, proof-pack.json, and a checksummed evidence manifest.
Inspect:
Boundary: Machine-readable files describe public facts only; explicit "do not infer" sections guard against overreach.