Pranay Saha memoji

Hey, I'm Pranay 👋

Computer Vision & ML Engineer

Building scalable AI inference systems & city-scale surveillance platforms.
Currently at Masterworks, Hyderabad.

Scroll
01 — About me

Turning pixels
into intelligence

I'm a Computer Vision & Machine Learning Engineer with a deep focus on building production-grade AI systems that operate at scale. From managing 4,000+ camera feeds to reducing computational overhead, I engineer solutions that are both technically rigorous and practically impactful.

🎓
Education
B.Tech Computer Science(AI)

Tech Stack

Python PyTorch TensorFlow OpenCV Docker Kubernetes Kafka FastAPI Git GitHub NVIDIA ONNX
02 — Experience

Where I've
made an impact

Masterworks
Computer Vision Engineer · SDE-1
Mar 2025 – Present Hyderabad, India
  • Contributed to the flagship low-code platform powering city-scale surveillance across highways, parks, and streets with automated incident detection and real-time alerting.
  • Architected multiple microservices enabling seamless one-click deployment of CV use cases across distributed edge devices.
  • Managed high-throughput model inference using Triton Inference Server, supporting 24×7 concurrent camera streaming across large deployments.
  • Built a dynamic multi-accelerator AI inference orchestration service converting models into optimised engines for NVIDIA and Qualcomm hardware.
  • Designed inference architecture processing real-time video feeds from 4,000+ cameras across heterogeneous edge devices.
  • Optimised end-to-end inference pipeline, achieving ~70% reduction in computational overhead.
  • Built user-configurable CV Use Nodes (Detection, Annotation, Classification, Alerting) enabling rapid workflow customisation.
  • Engineered data pipelines and trained CV models for city surveillance tasks including entry/exit monitoring and access alerting.
TritonTensorRTONNXDockerKubernetesKafka
Yukin
Machine Learning Engineer
Feb 2024 – Dec 2024 Remote
  • Worked on building a POC for a Digital Avatar solution using lip-sync with controlled facial expressions and voice cloning, mainly used for training and onboarding content across industries
  • Experimented with open-source lip-sync models like Wav2Lip and improved the inference pipeline by introducing batching and caching. This reduced memory usage by 50% and made it possible to process longer videos smoothly..
  • Evaluated multiple TTS and voice cloning models, including BARK and ElevenLabs, based on audio quality and inference time to select optimal solutions.
  • Built AI-driven content generation tools using GPT-3.5 Turbo to create personalised scripts, integrating database queries and external APIs for contextual content.
  • Collaborated with the Business team to deliver demo solutions for use cases such as client onboarding, employee training and personalised customer engagement.
FastAPITTSVoice CloningPython
Proglint Solutions
Computer Vision Engineer
Contract / POC
  • Constructed a Proof of Concept (POC) for real-time person Re-identification and multi-camera tracking, designed to enhance surveillance coverage in retail environments.
  • Engineered a highly precise model boasting a 98% precision rate, significantly slashing False Positives by 30%.
  • Streamlined algorithmic performance for enhanced speed and efficiency, reducing frame computation by 50%, leveraging advanced techniques such as multi-threading for seamless parallel processing.
Re-IDMulti-Camera TrackingOpenCVMulti-threadingPython
03 — Blog

Things I write
about

Triton Inference Server
Inference · MLOps 5 min read

Why Companies End Up Using Triton Inference Server — A Simple Case Study

A deep-dive into why Triton Inference Server has become the go-to choice for production ML inference — exploring real bottlenecks, benchmarks, and the engineering trade-offs companies face when scaling model serving.

Pranay Saha →
AI Model Formats
LLMs · Optimization 7 min read

How Model Format Alone Can Cut Your AI Infrastructure Cost by 70%

Learn how choosing the right model format can dramatically reduce inference costs, memory consumption, and deployment complexity without changing the underlying model architecture.

Pranay Saha →
AI Model Formats
AI · Multimodal 8 min read

Transforming Video into Knowledge — High-Traceability AI Learning Pipeline

In the EdTech and Learning Management System (LMS) sectors, video content is king. However, traditional video has a fundamental limitation: it is inherently unstructured, linear, and difficult to navigate.

Pranay Saha →
Read all on Medium →
04 — Contact

Let's build
something together

Whether it's an exciting AI opportunity, a collaboration, or just a chat about inference systems — my inbox is always open.