Praveen Kumar

Global logistics technology platform powering freight forwarding and supply chain operations.

LLM post-training: hackathon prototype → taxonomy agent → production evaluation

• Won Flexport’s 2026 Summer Hackathon by building a reusable QLoRA training and evaluation pipeline across tool-calling operations, document generation, and natural-language-to-Python workloads.
• Extended the approach to invoice taxonomy classification, fine-tuning Qwen3.5-9B on AWS GPU instances using historical operational data and iterating on dataset quality, training configuration, and agent-context alignment.
• Implemented GRPO on SFT-trained LoRA adapters, sampling eight responses per prompt and designing rewards around correct automation, human review, and costly high-confidence errors.
• Evaluated SFT and RL checkpoints within the full agent on 300 cases: accuracy increased from 66.7% to 69.0%, but errors among high-confidence predictions rose from 10.3% to 16.1%, revealing a regression in reliability and modeled business value.
• Diagnosed weak GRPO learning signals: 64.5% of sampled groups in an initial 4B experiment had identical rewards, yielding zero group-relative advantage.
• Benchmarked vLLM configurations and concurrency; collaborated with Commerce Engineering and Baseten on deployment and evaluation of the fine-tuned model for inference, comparing throughput, latency, and GPU-hosting economics against API inference at observed workload volumes.

Read the technical write-up ↗

SortBot: Autonomous warehouse pick-and-place robot

• CAD-designed and 3D-printed a 6-DOF dual-arm (SO-101) sorting station with vision-based perception; trained an imitation-learning policy (ACT/LeRobot) from teleoperated demonstrations.
• Integrated with warehouse systems and tested in a live warehouse environment in Wisconsin: a combined hardware + software product applying my mechanical engineering background.

Watch the warehouse demo ↗

Port Visibility: video understanding for port operations

• Goal: Capture and identify every auditable container-transfer event in ports, report a count range with per-event evidence, timestamps, and uncertainty, and abstain when footage does not support a reliable conclusion.
• Built the capture pipeline: evaluated viewpoints in Seattle and Tacoma, captured container-ship footage (phone, USB webcam, DSLR), and automated Mac-controlled timestamped clip capture.
• Ran systematic experiments across open VLMs (Gemma, Molmo, Qwen, InternVL, VideoLLaMA, TimeLens) and one-/multi-shot prompting. Built the strongest pipeline as structured CV + staged verification: stabilization, fence masking, contrast enhancement, motion detection, object detection and tracking and crane-geometry rules.

Engineering delivery (core role)

• Built and shipped BoomiCodeReviewAssistant to production: an AI-powered code review platform (Python, Claude) automating PR-level review of Boomi integrations: parses XML into architecture graphs, flags bugs, dependency risks, and security deviations. Adopted by the Ecosystem team for 200+ production reviews; migrated to AWS.
• Designed and built multiple highly scalable connectors in FUSE (data-ingestor): NetSuite, CSV, Excel, SFTP, and API. Built worker-management features including scaling, downloads, account management, and authentication.
• Led an AWS security remediation initiative: migrated the Ecosystem team from long-lived shared IAM credentials to Okta SSO-based temporary sessions for S3, and Boomi's AWS integration from IAM users to IAM Roles Anywhere with X.509 certificate-based authentication reducing exposure from long-lived credentials, aligned with zero-trust best practices.
• Led cross-functional collaboration across 4 engineering teams (Seattle, India, China) to ship the multiple inventory use cases on a tight timeline.

Leading global provider of online travel services.

• Led the requirements gathering and development of QCE Test Data as a Service in Java, enabling internal teams to simulate customer workflows such as account creation and Identity Management, loyalty management, and bookings.
• Reduced test account setup time from 5 hours to 5 minutes and booking creation time from 4 hours to 10 minutes.
• Evangelized best security practices around identity management and gave talks within the company on security concepts like SSO, OAuth, Okta, Active Directory and SAML.
• Built Data Lake infrastructure for analytics monitoring which surfaced through Tableau.
• Marketed QCE Test Data as a Service to new customers within Expedia increasing adoption from 1 team to 6 with number of accounts increasing from 60 accounts to 1500 accounts.
• Advocated for quality and reliability by establishing a QCE Test Data Dashboard using React and Python.
• Surfaced actionable metrics for leadership, enabling data-driven decision-making and reducing deployment risks by ensuring quality benchmarks.
• Regularly troubleshoot issues across APIs (Datadog), logs (Splunk), and infrastructure (Kubernetes), identifying and resolving bugs affecting customer flows.

Start-up that delivers tech-powered, omnichannel logistics and warehousing for e-commerce and retail partners.

• Early employee; built scalable integration systems for enterprise logistics clients.
• Built EDI Workflow Services in Scala and TypeScript to replace Dell Boomi integrations, streamlining third-party data pipelines for Fortune 500 partners — driving a 50% increase in user load, 25% reduction in integration time, saving $400K annually, and reducing bugs by 20%.
• Designed and launched a React-powered API documentation portal, reducing partner onboarding time by 25%, and served as the go-to technical contact for Fortune 500 partners on APIs, webhooks, and security.
• Upgraded Node.js from v10 to v18, componentized frontend to Storybook, and implemented feature flagging infrastructure (LaunchDarkly, FeatureHub) to accelerate release cycles.

Leading global provider of online travel services.

• Led migration of on-prem services and databases to AWS — a multi-team, multi-region, multiyear effort collaborating with solution architects, engineers, legal, and Amazon, resulting in the deprecation of hundreds of datacenter servers.
• Owned GDPR compliance strategy for European services, working with legal and regulatory teams to design data deletion architecture — addressing an estimated $2M in potential compliance exposure.
• Implemented Elasticsearch as a fallback to Solr/Cassandra clusters, reducing latency by 20% during throttling, enabling new post-booking revenue features, and rewriting onboarding code with Spring Integration and Swagger to quickly add new APIs.
• Partnered with Bank of America engineering teams to onboard them to the Expedia Network, including SSO authentication, API integration, and third-party audit infrastructure.

Leading computational platform used across multiple domains.

• Developed and maintained the vision, strategy, and roadmap for executing test plans using Mathematica and C++.
• Improved test coverage by 45% by creating new test suite in Python and then visualizing all quality metrics on internal dashboard
• Collaborated with software engineering teams to drive the initial Wolfram Alpha project to completion. Was part of the launch committee, responsible for the analytics dashboard and the customer feedback platform.