Trusted by AI-first companies worldwide →
ML Platforms & Kubernetes
I build Kubernetes-native ML platforms — custom resources, autoscaling GPU inference, and SageMaker-equivalent pipeline tooling — that cut cost and latency without sacrificing reliability or developer experience.
LLM Orchestration & GenAI
I design centralised LLM gateways with async request routing, safety guardrails, and observability so that every team in an organisation can consume models through one governed, well-instrumented API.
Agentic Systems
I build agentic platforms that take on real engineering work — migrating legacy inference stacks with zero production impact, triaging issues via tag-driven CI/CD, and enforcing production best practices automatically.
MLOps & Inference Infrastructure
I ship production MLOps pipelines with Docker, vLLM, experiment tracking, model versioning, INT8 quantization, and automated training/deployment — reducing deployment time and inference cost while holding accuracy.
Streaming & Data Pipelines
I architect real-time CDC and event-driven pipelines on AWS (Kinesis, Kafka, Flink, EventBridge, Lambda) with IaC via AWS CDK, processing billions of rows for live KPI dashboards and cross-account workflows.
Edge AI & Vision Systems
I optimise computer-vision models for the edge — YOLOv7/v8 and VideoMAE on custom Jetson SOMs with TensorRT and CUDNN — delivering sub-second analytics for thousands of devices in the field.
- Led an in-house Kubernetes orchestration platform delivering SageMaker-equivalent ML pipeline functionality via custom resources, achieving 30% resource reduction, 40% latency improvement, and 45% cost savings across GPU inference workloads.
- Designed a centralized LLM orchestration platform routing async request queues across downstream services with built-in safety guardrails and observability, giving every engineering team a unified, governed API for model consumption.
- Designed and shipped an Agentic System to migrate a legacy inference stack (50K+ lines of code) serving ~10K requests/sec with zero production impact, reducing annual GPU compute cost by ~$130K (~84%).
- Leveraged Claude and an in-house Agentic AI platform to map all internal agents across the monolithic codebase, enabling automated issue triage and resolution via tag-driven GitHub CI/CD workflows while enforcing production best practices.
- Deployed production-grade RESTful APIs using FastAPI and Python on AWS Lambda with automated CI/CD pipelines, integrating SonarQube and Checkmarx for code quality and unit/integration testing — handling 7M+ daily requests and cutting deployment time by 40%.
- Built end-to-end ML pipelines with MLOps practices using Docker and vLLM — experiment tracking, model versioning, automated training/deployment — reducing model deployment time from 2 hours to 30 minutes, with INT8 quantization shrinking model size by 75% at 95% accuracy.
- Designed and implemented a real-time CDC streaming pipeline from SQL Server to AWS Kinesis using Kafka, Flink, and AWS CDK (IaC) to process ~5 billion rows of structured data and feed real-time KPI dashboards for leadership decision-making.
- Architected a receipt classification pipeline using AWS Lambda, GenAI Platform, and FastAPI with Amazon Nova models, deployed across multiple AWS accounts via EventBridge for async cross-account orchestration — improving recognition accuracy by 16%.
- Implemented scalable, fault-tolerant video streaming infrastructure using MQTT, AWS Kinesis, REST APIs, RDS (PostgreSQL), and DynamoDB to process real-time feeds from 5000+ edge devices with sub-1s latency — achieving 99.9% uptime and reducing data transmission cost by 45%.
- Pioneered a real-time table tennis analytics system deployed at World Table Tennis events. Built an end-to-end ML pipeline using YOLOv7 for ball detection (99.6% MAE) and a fine-tuned VideoMAE (96% accuracy), delivering ball speed, pitch location, and shot classification with Sony Live broadcast integration.
- Optimized AI firmware for custom Jetson SOMs, achieving 70% performance boost and 38% space reduction while maintaining model accuracy within 2% for edge deployment of YOLOv8 using TensorRT and CUDNN.
| # | Client | Year | Tags | Domain | |
|---|---|---|---|---|---|
| 01 | Kubernetes ML Orchestration PlatformTractable | 2025-2026 | #kubernetes#gpu#mlops#sagemaker | ML Platform | → |
| 02 | Centralized LLM Orchestration GatewayTractable | 2025-2026 | #llm#observability#guardrails#async | GenAI | → |
| 03 | Agentic Legacy Inference MigrationTractable | 2025-2026 | #agents#claude#refactor#cost-savings | Agentic | → |
| 04 | Receipt Classification GenAI PipelineBasecone, Wolters Kluwer | 2024-2025 | #nova#lambda#eventbridge#fastapi | GenAI | → |
| 05 | Real-Time CDC Streaming PipelineBasecone, Wolters Kluwer | 2023-2024 | #kafka#flink#kinesis#aws-cdk | Data Platform | → |
| 06 | FastAPI Inference at 7M req/dayBasecone, Wolters Kluwer | 2023-2025 | #fastapi#lambda#vllm#mlops | Backend | → |
| 07 | Real-Time Table Tennis AnalyticsCamereye | 2021-2023 | #yolov7#videomae#sony-live#cv | Sports AI | → |
| 08 | Edge Video Streaming at 5000+ DevicesCamereye | 2019-2023 | #mqtt#kinesis#jetson#tensorrt | Edge AI | → |
“Rahul is the rare engineer who pairs deep ML expertise with the systems instincts to ship platforms — not just models. His work on our Kubernetes orchestration and LLM gateway reshaped how every team here consumes ML.
I'd put him on any team building serious AI infrastructure.”
“Rahul owned our most demanding ML services at Basecone — FastAPI on Lambda serving 7M+ daily requests, real-time CDC pipelines moving billions of rows, and a GenAI receipt classifier deployed across multiple AWS accounts.
He delivered with rigor on every front: code quality, MLOps, and architecture.”