I build the backend half of AI products — LLM agents that run in parallel, retrieval that cites its sources, and Python services that stay up. 5 years across financial services and autonomous vehicles, shipping production systems, not demos.
I'm a backend AI engineer. I build the unglamorous half of an AI product: Python/FastAPI services, LLM inference adapters, schema-validated outputs (Pydantic, typed parsers), retrieval that cites its sources, and the eval harnesses that stop a prompt change from quietly breaking production.
At a Fortune 100 Financial-Services Company I built parallel LLM agent workflows for incident intelligence — specialised agents reading operational data concurrently, merged into source-grounded summaries with confidence scoring — which cut complex-case resolution time 40%. On the reliability side I automated multi-region AWS failover for 32 enterprise applications, cutting per-app cutover time 70%, and validated RTO/RPO with Gremlin chaos experiments. Earlier, on autonomous-vehicle programs for a German automotive Tier-1 (ZF), I built ADAS perception models.
Outside work I ship the same ideas as open source — a RAG document assistant with hybrid retrieval and a recall@k eval harness, an LLM token-compression library, a token-aware inference gateway, a chaos-testing toolkit, a Kubernetes failover operator. I like AI that can be measured, and systems that survive a bad day in production.
Production LLM workflows and multi-region resilience in a regulated financial-services environment.
Deep-learning perception for autonomous-vehicle ADAS programs.
GPA 3.5 / 4.0 · Hoboken, NJ. Focus on ML systems and software engineering.
LLM features that ship: agents in parallel, retrieval that cites, outputs that validate.
Python services behind the models — typed, tested, observable.
Proof that a prompt or model change didn't quietly break anything.
Multi-region failover, chaos experiments, reproducible infrastructure.
Classical models to CV deep nets, trained and evaluated rigorously.
RAG service that answers questions over uploaded documents with citations — or says "I don't know" when the context doesn't cover it. Chunking, embeddings, hybrid keyword + vector retrieval, Claude generation behind FastAPI, and a recall@k eval harness so a retrieval change can be measured, not guessed at. Fully offline-testable with an echo LLM stub.
Isolation Forest flags the anomalous window, then three LLM agents run in parallel — root cause, blast radius, remediation — and merge into one source-grounded report. The open-source shape of the agent workflow I built at work.
A gateway in front of LLM providers: least-outstanding-tokens balancing (not round-robin), circuit breaking on unhealthy backends, semantic caching, and per-caller cost accounting exported to Prometheus.
Zero-dependency Python library that cuts LLM input-token costs 30–70% via a budget-driven compression pipeline — six strategies on an aggressiveness ladder, protected-region shielding, and a drop-in Anthropic SDK middleware.
Injects latency, error and blackhole faults through a reverse proxy, then scores the steady-state hypothesis with p50/p95 probes and a pass/fail resilience report. Stdlib-first, no mocks.
A FailoverApp CRD that reconciles its Deployment and reports readiness against a minHealthy replica floor — real health, not just "the object exists". Built with kubebuilder / controller-runtime.
Trains, versions and serves a credit-default model behind a schema-validated FastAPI service — seeded runs, no train/serve skew, versioned artifacts, and a CI matrix gating on ROC-AUC thresholds and the full API surface.
Ships a containerized service to AWS ECS Fargate behind an ALB with CloudWatch logs, alarms and a dashboard — modular networking/IAM/monitoring, delivered by GitHub Actions with keyless OIDC auth.
Verifies construction submittals against CSI spec sections automatically — flags missing items, standard mismatches and value deviations, turning a manual page-by-page review into a diffable report.