AI/ML Engineer (Research) • Agentic AI • Full-Stack • DevOps/MLOps • Distributed Systems • Computer Vision
I build agentic systems end-to-end: model + retrieval + evaluation, planner → executor loops, UX, and deployment. Reliability is how I ship AI safely — idempotency, retries/backoff, DLQ, reconciliation, audit logs, regression gates.
Proof
- Delivered 3× throughput while holding p95 ≤ 3.8s by pairing guardrailed agent flows with regression gates + evidence bundles.
- Tuned platform schedulers for ~50% less swap thrash and ~$40K/year infra savings via idempotent queues, retries/backoff, and DLQ hygiene.
- Cut 38% inference cost yet kept 60fps UI at 100k+ rows through caching/batching, virtualization, and observability-led tuning.
Background: NYU MS CS (May 2026) • Samsung Research • Veach AI • Research Assistant • ES 2026 full paper accepted.
- Agentic AI systems – planner → executor loops, tool use, eval harnesses, safety/guardrails, and logs that keep auditors happy.
- ML/AI pipelines – RAG + LLM apps, training/eval harnesses, latency/cost optimization, experiment tracking.
- Computer Vision – OpenCV/vision-model pipelines, dataset tooling, scoring dashboards.
- Full-stack product – Next.js/React frontends, FastAPI/REST backends, Postgres/Redis data layers.
- DevOps/MLOps – Docker/Kubernetes, CI/CD (GitHub Actions), observability, deployment runbooks.
- Event-driven workflow orchestrator — workflow-orchestrator-sandbox. Agent + ML pipeline engine (FastAPI + Redis + Postgres) with idempotency keys, retries/backoff, DLQ, reconciliation sweeps, and audit-ready metrics. Demo/Docs: repo README.
- RAG evaluation + latency/cost harness — rag-eval-harness. Deterministic dataset loader, caching vector store, async workers, latency/cost dashboards, and CI-ready eval harnesses that drove the 38% cost win. Demo/Docs: repo README.
- High-volume analytics UI — Portfolio. Full-stack (Next.js + APIs + DB) analytics experience with virtualization, workerized transforms, and instrumentation to keep 60fps at 100k+ rows. Demo/Docs: https://karan-allagh.vercel.app.
- Low-latency C++ prototyping — Samsung/Veach internal (non-public). Near-metal agentic kernels for SIMD batching, pipeline hazard detection, and telemetry to hold p95 ≤ 3.8s. Demo/Docs: available under NDA.
- Cloud automation agent — Cloud_Automation_Agent-. Electron + Django + Orion agent stack that plans, executes, captures evidence, and enforces guardrails for cloud operations. Demo/Docs: repo README (video WIP).
- Bake reliability patterns (idempotency, retries/backoff, DLQ, reconciliation, audit logs) into every agentic or ML system to keep rollouts safe.
- Reliability is not a phase; it’s the guardrail for AI/ML features before they reach customers.
- Languages: Python, TypeScript/JavaScript, Java (Spring), C++17/20, Go, SQL, Bash.
- ML/AI: LLM tooling (agentic planners, RAG, eval harnesses), cost/latency tuning, dataset tooling.
- Computer vision: OpenCV + vision-model pipelines, dataset scoring + regression dashboards.
- Web & APIs: React/Next.js frontends, FastAPI/REST services, Postgres + Redis data layers.
- Infra / DevOps / MLOps: Docker, Kubernetes, GitHub Actions, Kafka streams, observability (logs/metrics/traces), runbooks + on-call.
- ES 2026 accepted full paper – Agentic Decomposition for Reliable Long-Horizon AI Planning (public preprint coming soon).
- Interest areas: agentic decomposition, evaluation rigor, retrieval quality, latency/cost trade-offs for LLM + CV workloads.
New Grad roles in AI/ML engineering, agentic AI systems, or backend/platform engineering — including Founding Engineer (0→1) opportunities at early-stage startups (NYC hybrid or remote). I love building reliable, observable systems: idempotency, evals, CI, and production-ready deployments.
📫 ka3527@nyu.edu • LinkedIn • Portfolio • GitHub
workflow-orchestrator-sandbox— shows idempotency, retries/backoff, DLQ, reconciliation.rag-eval-harness— demonstrates latency/cost benchmarking and async evaluation.Cloud_Automation_Agent-— agentic automation with plan reviews + audit logs.Portfolio— high-volume UI + recruiter-ready story.office-submission— reliability patterns inside Office add-ins.