Maaz Karim · vibe

vibe

small fragments of life & coding.

Not formal enough for blog posts, but still part of the story.

shipping Remote #VideoVerse#News#Reliability#OCR

VideoVerse · news production ML

Later the work moved deeper into news pipelines, where the focus shifted toward production reliability, OCR, and publishing quality under live traffic.

  • tightened live stream reliability with better retries, recovery paths, and safer concurrency
  • shipped a GPU OCR service on Ray Serve with warmup, autoscaling, and lower-latency inference
  • improved metadata extraction and routing for automated publishing quality
  • added tracing, structured logs, and stronger tests across the stack
deep work India · Remote #VideoVerse#Sports#FastAPI

VideoVerse · sports video systems

This phase was mostly about sports video systems, where the job was turning researching media pipelines into services teams could actually rely on.

  • refactored a live video pipeline toward a more modular, async-first design
  • built multilingual audio dubbing services for sports workflows with fault-tolerant FastAPI backends
  • helped in finetuning classification models for event detection using VidLLMs using InternVL-2
  • added practical MLOps pieces like MLflow, LiteLLM, Datadog, and Sentry. Blog on MLOps@Videoverse
building Bengaluru · Hybrid #AI Agents#Platform#MLOps

TrueFoundry · platform and tooling

This phase pushed me closer to platform engineering and developer tooling.

  • built and extended an AI agent CLI for deploying services and jobs
  • shipped POC for AutoDeploy, aimed at reducing deployment setup friction
  • added support for Locust-based load testing LLM benchmark
  • helped move internal services from Flask to FastAPI and tightened CI, docs, and maintainability
iterating Bengaluru · Remote #Dive#Meetings#LLMs

Dive · meeting AI

After the Octernship, the work extended into Dive itself, contributed in building product.

  • built meeting summarization, chapterization, and contextual retrieval prototypes
  • improved transcript-based summary quality with evaluation tooling
  • optimized parts of the stack for faster real-time LLM inference
shipping Remote #GitHub#Octernship#NLP

GitHub Octernship

Worked around NLP and ML systems.

  • worked on transcript summarization pipelines and internal evaluation tooling
  • improved relevance for Ask-AI style summaries through tighter iteration loops
  • contributed across audio and NLP focused ML tasks during the program
hands-on Bengaluru · Hybrid #Medbikri#OCR#Retrieval

Medbikri · document ML

Worked around Document OCR for medical bills

  • improved medical search quality with Elasticsearch query logic and filters
  • built pre-processing pipelines for bills using Pix2Pix, CGAN, and YOLOv8
  • fine-tuned Donut for structured bill extraction into PostgreSQL
  • combined embeddings with retrieval to make extraction and lookup more useful
learning fast Remote #Fellowship.AI#Data Science#Modeling

Fellowship.AI · early data science

An early research-heavy internship where the focus was prediction quality under incomplete real-world data.

  • modeled campaign and cohort behavior for better digital marketing recommendations
  • generated large synthetic datasets to test cohort-aware approaches
  • explored ways to recover cohort structure when it was missing from production data