Oscar Ndugbu
Staff Backend and Platform Engineer
Backend, platform, and AI infrastructure built for visible failure, clean recovery, and decisive operational handoffs.
Operating profile
UTC+1
Staff Backend and Platform Engineer
Decision-led engineering for systems that must remain observable, recoverable, and understandable under pressure.
Focus
Distributed systems · platform reliability
Proof
TaxBridge · SabiScore · SwarmXQ
Method
Constraint → decision → outcome → evidence
The work is shown as a decision record: what the system had to survive, what was chosen, what changed, and where the evidence lives.
Tax workflow orchestration with idempotent replay, database-enforced tenant isolation, and an explicitly private evidence record.
TaxBridge Reliability Ledger
Constraint
Rate-limited tax submission during deadline bursts, with duplicate filing risk on retry.
Decision
Database RLS plus idempotent BullMQ jobs keyed to submission content.
Outcome
Retries can resume without re-submitting a completed filing operation.
Evidence
Private architecture record · source available to verified employers.
Tax workflow automation for Nigerian small businesses. PostgreSQL RLS isolates each tenant at the database level. BullMQ jobs use submission-derived idempotency keys to prevent duplicate processing through retries, and the audit trail is hash-chained. NRS DigiTax integration remains in progress.
Ensemble ML serving with versioned caching, model-quality telemetry, and a deliberate lower-confidence fallback path.
SabiScore Reliability Ledger
Constraint
Inference and model-health signals must remain usable through cache or feature-store failure.
Decision
FastAPI serving, versioned Redis caching, Pub/Sub fan-out, and a lower-confidence fallback model.
Outcome
Dependency failure returns a labelled lower-confidence result instead of stale output or an opaque error.
Evidence
Public source · architecture case study; performance artifacts are not published.
Multi-agent orchestration with task-specific model routing, bounded fallbacks, checkpoint recovery, and an operator dashboard.
SwarmXQ Reliability Ledger
Constraint
Local inference must stay correct under constrained memory and intermittent third-party connectivity.
Decision
Task-specific model routing with checkpointed state and bounded fallbacks.
Outcome
Failed work resumes from the last consistent checkpoint instead of restarting the run.
Evidence
Public source · architecture case study.
ZK proofs for document integrity verification — confidentiality and verifiability as simultaneous properties, not a tradeoff.
Hashablanca Reliability Ledger
Constraint
Verify document integrity without exposing the document or exceeding verifier gas budgets.
Decision
Groth16 proofs generated off-chain with per-network verification adapters.
Outcome
Integrity can be checked without revealing document contents.
Evidence
Public source · architecture case study.
Federal education reporting across 36 state sources — probabilistic deduplication, per-state retry semantics, and explicit partial-output handling.
UBEC Data Pipeline Reliability Ledger
Constraint
Late and partial state submissions must not block every completed report.
Decision
Isolated Airflow task groups per state, followed by cross-state validation gates.
Outcome
Thirty-five state reports can proceed while one delayed source remains visibly pending.
Evidence
Private federal engagement · architecture case study.
The evidence is strongest when the failure boundary is named: what can break, how the system responds, and which record supports the claim.
COMPLIANCE PLATFORM · FINTECH
Duplicate filing and tenant leakage are designed out of the critical path.
BullMQ absorbs NRS API rate limits with idempotent submission keys. PostgreSQL RLS keeps tenant isolation at the database engine, while the case study records the integration status and private evidence path.
Chosen: RLS at engine level, not application-layer filtering — NRS audit scrutiny demands proof that tenant data cannot cross-contaminate.
ML PLATFORM · OBSERVABILITY
Inference stays useful when cache or feature-store dependencies fail.
Versioned Redis caching and a lower-confidence baseline model provide a deliberate degraded path. The measured record is bounded to a 90-day Prometheus window and a named reactive-alerting baseline.
Chosen: FastAPI + Redis Pub/Sub — sub-50ms event fan-out at sustained load, impossible with synchronous polling under concurrent sessions.
AI AGENT PLATFORM · ORCHESTRATION
Agent work resumes from a known state instead of replaying the entire run.
Task-specific model routing operates within an 8GB local-inference budget. Checkpointed state, timeout budgets, and bounded fallbacks make partial failure visible and recoverable.
Chosen: Autonomous evolution layer over static agent configs — manual tuning cannot adapt to novel inputs at scale.
FEDERAL INFRASTRUCTURE · DATA
One late state submission does not block every completed report.
Isolated Airflow task groups, canonical ingest schemas, probabilistic record linkage, and cross-state validation gates keep partial output explicit across 36 state sources.
Chosen: Probabilistic deduplication over exact-match — state submissions use inconsistent school name spellings across ministries.
Four source-linked packages and experiments focused on tenant isolation, auditability, incident triage, and local-model routing. Each card names the evidence path it can actually support.
The evidence contract is simple: public-source claims link to public source; private project context is not presented as third-party verification.
Node.js · PostgreSQL
Public sourceA PostgreSQL Row-Level Security helper for moving tenant-isolation policy closer to the database boundary instead of relying only on application checks. The repository documents the implementation and its intended use in multi-tenant systems.
Fintech · Auditability
Public sourceA hash-linked audit-log pattern designed to make retrospective record changes detectable during integrity checks. The public repository exposes the implementation so the chain construction and verification approach can be inspected directly.
AI · DevOps
Public sourceAn incident-triage experiment that turns logs and traces into ranked root-cause suggestions for operator review. Its public source makes the input path and ranking workflow inspectable without presenting an unmeasured time-savings claim.
AI · Agent Orchestration
Source-linkedA local-model routing pattern derived from SwarmXQ. It assigns task classes to different GGUF models and defines fallback chains, with the linked public SwarmXQ source providing the inspectable evidence path for the routing approach.
Public evidence paths are available on GitHub.
The stack is organized by operational responsibility: isolation, observability, recovery, data movement, and delivery. Each tool maps to a recorded system decision rather than a generic proficiency list.
Isolation
data boundary
TaxBridge · Fastify 5 · PostgreSQL 15 RLS
Telemetry
operating signal
SabiScore · XGBoost · LightGBM · Prometheus
Recovery
failure path
SwarmXQ · Ollama GGUF · Checkpoint Recovery
Core production stack
Default view is intentionally curated for fast trust. Open the full explorer only when you want deep implementation coverage across all 62 tools.
+6 more in the full explorer below.
Production system × skill lineage
Full explorer · 62 tools · 8 pillars
Expanded layer for CTO and engineer-level scrutiny. Optional by design to preserve reading momentum for recruiters and founders.
More than a decade in federal data systems, including reporting pipelines across 36 state sources, followed by backend, platform, and AI infrastructure work built around failure-aware delivery.
The UBEC work centered on heterogeneous submissions, partial state data, and reporting that could not silently accept bad inputs. That operating context carried into every system after it: explicit failure modes and architecture decisions written so the next engineer can understand the reasoning long after deploy.
TaxBridge: React Native / Expo 54 mobile app, Fastify 5 API, and PostgreSQL 15 RLS for database-enforced tenant isolation. SabiScore: ensemble ML inference behind a Next.js dashboard with caching, telemetry, and a deterministic fallback path. SwarmXQ: checkpointed agent orchestration with explicit retry and recovery semantics.
The current resume records the professional history and certifications; public repositories provide the inspectable layer. The work is the record, and private evidence is labelled as private.
Constraint Code · Non-Negotiable Standards
Correctness is a product feature, not a backend preference.
The database schema, API contracts, and UI flows all carry the burden of truth — not just the last layer to touch the request.
Silent failures are design failures.
Retries, dead-letter handling, structured logs, and metrics exist before the first incident, not after the apology.
Every critical decision deserves a written rationale.
The next engineer should be able to understand the choice, the tradeoff, and what was consciously rejected.
If a system cannot be observed, it cannot be trusted.
Health checks, traces, and dashboards are treated as part of the product surface — not optional operations decoration.
Traced to architecture decisions across TaxBridge · SabiScore · SwarmXQ
Based in Lagos. Operating on UTC+1.
Architecture, ML systems, failure modes, and the decisions behind them — written for engineers who need the reasoning, not just the result.
Start here if you want to understand how these systems were designed.
How variable power, network, and regulatory conditions shape more resilient infrastructure decisions. Constraint is treated as a design input, not a branding claim.
Have a system problem worth examining? Start with the stakes →
For Staff backend and platform roles, architecture reviews, AI infrastructure, or reliability work: send the problem, the stakes, and the timeline.
Systems that work at 2am.
That's the standard.