upavit.
Resume

WORK / SELECTED SYSTEMS

Selected work.

My work at Gradient has included robotics, backend services, data pipelines, GPU inference, and deployment infrastructure. Here are some of the projects and problems I’ve worked on.

GPU inference and benchmarking

Model serving, resource allocation, and performance testing for GPU-backed services.

WHAT CHANGED

  • Repartitioned an 8-GPU vision-language plane across four model configurations using measured memory and throughput behavior.
  • Corrected a benchmark that overstated throughput by roughly 4.6× when repeated inputs hit caches.
  • Built a GPU vector pipeline recorded at about 33,480 items an hour with both GPUs fully utilized.

TECHNOLOGIES

PythonvLLMTritonKubernetes

Search and data pipelines

Built and maintained the data paths behind a collectibles platform with a primary corpus of roughly 21–22 million records.

WHAT CHANGED

  • Connected PostgreSQL to search and vector stores through Debezium and Kafka, with monitoring, schema evolution, backfills, and recovery tooling.
  • Reworked a large-table backfill to scan bounded page windows. The estimated PostgreSQL planner cost per batch fell from 7,079,081 to 53,971—about 131× lower, a planning estimate rather than a measured speedup.
  • Preserved bounded writes and lock timeouts so background backfills could coexist with application work.
  • Built operator workflows for reviewing staged changes, making bulk updates, and recovering interrupted data operations.

TECHNOLOGIES

PostgreSQLDebeziumKafkaElasticsearchMilvusTypeScriptPython

Cloud and on-prem infrastructure

Infrastructure designed to make networking, deployment, and recovery explicit across cloud and on-prem systems.

WHAT CHANGED

  • Built shared AWS foundations for private discovery, GitHub OIDC, VPC networking, and separately managed Tailscale routing.
  • Moved a pricing service through a blue/green RDS cutover and added guarded Debezium-to-Kafka CDC.
  • Operated hybrid GPU and Kubernetes environments with Terraform, Ansible, Kustomize, Docker, and observability tooling.
  • Turned recurring deployment assumptions into reviewed infrastructure and configuration.

TECHNOLOGIES

AWSTerraform/OpenTofuKubernetesTailscaleRDSCloudWatchSecrets ManagerPrometheus

Payment retries and recovery

Commerce flows built to recover when a provider, worker, or network leaves the outcome uncertain.

WHAT CHANGED

  • Made marketplace outbox work deploy-safe so worker replacement could not leave paid jobs permanently claimed.
  • Hardened cashouts with idempotency, reconciliation, and escalation for ambiguous provider outcomes.
  • Modeled instant payout as its own irreversible money movement.
  • Kept fulfillment and seller notifications outside source-of-truth transactions with signed, idempotent delivery.

TECHNOLOGIES

PostgreSQLStripe ConnectOdooOutbox patternsIdempotencyBackground workersTerraform

Internal operations tools

Applications for provisioning jobs, managing inventory, reviewing changes, and monitoring machines.

WHAT CHANGED

  • Built the application shell and expanded it with provisioning, orders, machine visibility, audit, search, and configuration workflows.
  • Delivered staged review, bulk update, scan-first inventory, requeue, and image-submission flows across APIs, storage, and jobs.
  • Added machine throughput, status, and version visibility for physical and software systems.
  • Made review, recovery, and feedback explicit for actions that change production state.

TECHNOLOGIES

React/Next.jsTypeScriptFastAPIPostgreSQLS3PrometheusAuth/RBAC

Camera capture and card handling

Software connecting cameras, scanners, and pneumatic card handling to the digitization pipeline.

WHAT CHANGED

  • Integrated camera capture and scanner acquisition with card identity and upload workflows.
  • Coordinated capture, upload, and machine commands with threads, queues, and controller locks.
  • Added vacuum and pressure checks, camera-timeout fallback, and reconnect behavior for interrupted operations.
  • Built operator controls for capture, pause, and recovery, and integrated camera configuration and validation into the workflow.

TECHNOLOGIES

PythonC++Cameras / scannersTWAINPneumatic handlingThreads / queues

Robotics and machine control

Control software for automated card sorting, including motion, image capture, and fault recovery.

WHAT CHANGED

  • Built host orchestration and embedded control for motion, homing, vacuum, pressure, and fault handling.
  • Built controller communication for movement, homing, status, and fault checks over Ethernet and Modbus.
  • Implemented route ordering for sorter visits using Held–Karp dynamic programming.
  • Later hardened firmware around motor faults, fail-safe stopping, and configurable rescan behavior.

TECHNOLOGIES

PythonC++Ethernet/TCPModbusTWAINOpenCVThreads/queuesEmbedded control