Hello, I'm Prassanna

Machine Learning Engineer

Solving problems using machine learning across the {data, modelling, infrastructure} stack.

01. About

I am a Machine Learning Engineer focused on the intersection of infrastructure, data, and modelling. My work revolves around building robust systems that enable the "Agentic Web," where AI agents can reliably operate, communicate, and execute complex workflows.

Currently, I'm simplifying the unification of Code and AI, designing protocols that allow agents to be more than just chat bots.

02. Stack

Languages

PythonTerraform

ML Frameworks & Serving

PyTorchFastAPILangChainAgent SDKsOpenTelemetryMCPONNX

Orchestration & Data

SkyPilotAWS/GCPTraining Job OrchestrationApache BeamDagster/AirflowTemporal

Compute & Infrastructure

KubernetesDocker

03. Experience

Principal MLOps Engineer + Platforms Team Lead

Papercup AI
Jan 2023 - PresentLondon, U.K.
  • Designed and led the ML cloud training stack from scratch with ClearML and Skypilot, scaling compute from 16 to 50 GPUs.
  • Led an initiative to migrate the model serving stack to be serverless, reducing idle cost time by 90%. Built an internal library to create consistent interfaces, conventions and OpenTelemetry tracing.
  • Designed and led an initiative to increase data processing speeds by 50x using Apache Beam with GCP's Dataflow using 250+ GPUs.
  • Implemented a tool using GPT-4 and LangChain to automate creative text generation, reducing data acquisition time from ~4 months to ~2 weeks.
  • Automated and productionized a transcription and translation workflow using Deepgram, GPT-4, LangChain, Temporal, FastAPI and Streamlit, allowing the asynchronous system to serve 1000% more workflows per day.
  • Created the foundational Platforms team (SREs, MLOps, Data Engineers) and hired two MLOps engineers to the team.

Senior Machine Learning Engineer

Zenith AI (Acquired by Opentrons)
Oct 2020 - Jan 2023Belfast, U.K.
  • Founding engineering team member, company acquired after 9 months.
  • Reduced cumulative QA build time from 2800 minutes to 200 minutes per week by optimizing Docker builds.
  • Designed an async queue-based graph executor reducing ML load time from 3 hours to 35 minutes.
  • Led a team implementing secure self-service ML web app deployments, cutting wait times from 3 days to 15 minutes.

Applied Computer Vision Researcher

AnyVision (Now Oosto)
Sep 2019 - Oct 2020Belfast, U.K.
  • Developed a Focus of Attention system for retail using 3D geometry and depth estimation.
  • Built a defect classification prototype in 48 hours, later piloted by a major electrical company.
  • Introduced a benchmark-driven-development research philosophy with Grafana analytics.

Image Processing Consultant @ HP

Altran (Acquired by Capgemini)
Mar 2017 - Sep 2019Barcelona, Spain
  • Supported the creation of a Python-based middleware for large format printers.
  • Implemented sensor-based calibration routines for large format and 3D printers.
  • Built internal tools to identify and fix calibration bugs, halving debugging time for R&D.

04. Education

PhD Student and Teaching Assistant

May 2013 - Feb 2017

Computer Vision Center

Implemented Slither, a random forest framework for semantic segmentation using C++ and Python.

MSc. Computer Vision

Sep 2011 - Sep 2012

University of Sheffield

Worked on face recognition, traffic analysis, and FPGA-based cryptography.

B.E. in Electronics and Communication

Aug 2007 - May 2011

Manipal Institute of Technology

Internships at FH Westkueste, Germany (2011) and mdirect, Tunisia (2009).