§00 · helloBENGALURU, INDIA · 12.97°N 77.59°Ehe/him

Hi, I’m Mark

ML Engineer building MLOps pipelines & systemsat 314e Corp

Mark, sitting at a café table, chin resting on his hand — me having fun

I'm an ML Engineer at 314e Corp in Bengaluru, where I build and operate MLOps pipelines for healthcare ML systems — orchestration, distributed data processing, and model serving at scale. Outside of work I write low-level Rust and Python for fun: reimplementing NumPy from scratch, and building terminal tools I actually use every day.

§01

Experience

git log --graph

314e Corp current

Bengaluru, India

(HEAD) Software Development EngineerJan 2025 – Present
Associate Software Development EngineerJul 2024 – Jan 2025

Owned the ML stack for a clinical document-processing product across 26 entity types, fine-tuning a 27B vision-language model to 96 percent test accuracy in production.
Led development of both the entity extraction flow and the agentic feedback flow, working with cross-functional teams to align requirements, interfaces and rollout so the two shipped as one connected system, while mentoring 2 junior software engineers.
Wrote a 7,600-line fine-tuning package (SkyPilot, Temporal, ClearML) with one dispatch layer, so Gemma 3, Qwen3.5-9B and Qwen3.6-27B train through the same code.
Built a reviewer-feedback retraining loop with stratified datasets and automated deploy behind a per-entity F1 gate, so no regression beyond a tolerance scaled to its prior score ships.
Built a weekly drift monitor where an LLM agent reads page OCR through a read-only MCP server, so regression is separated from labelling change by checking the document itself.
Wrote a schema-driven extraction layer with per-field token-logprob confidence, giving 4 LLM providers and 3 inference engines one validated output contract.
Trained a calibrated XGBoost model on logprobs, OCR grounding and embeddings over 99,427 reviews, halving false approvals (1.4% vs 2.7%), cutting calibration error 0.139 to 0.004 and retiring 83% of manual entity review.
Moved serving from Transformers to vLLM to SGLang on RunPod serverless with baked weights and pre-warmed kernels, reaching 64 concurrent jobs per worker with no cold-start cost.
Switched training to map-style lazy loading with VRAM-sized batches after a cap bounded sample count, not bytes, exhausted 80 GB of host RAM, cutting training from 90 to 25 hours.

Engineering writeups of this platform, which I contributed to — published by 314e, authored by Dr. Srivatsan Sridhar: AI document extraction & classification · IDP for medical records

Banach Technologies

Machine Learning Intern

Singapore (Remote)
Apr 2024 – Jul 2024

Built and deployed ML models against live market APIs to execute hedging strategies, cutting data-processing latency through targeted performance tuning and covering the core trading libraries with unit tests.

Alemeno

Software and ML Intern

Maharashtra (Remote)
Jul 2023 – Apr 2024

Built a preprocessing pipeline for semantic-segmentation models over large-scale GIS raster data, implementing distributed computer-vision algorithms including Douglas–Peucker and Jarvis March, and deployed Django services on AWS behind end-to-end CI/CD.

Bewgle

Data Analyst Intern

Bengaluru (Remote)
May 2022 – Jul 2022

Applied NLP to proprietary Amazon review datasets for product-trend insight, automating preprocessing in Python and Bash to cut data redundancy 90%.

§02

Publication

IEEE Xplore · 10119431

cited ×25

Medical Waste Classification using Deep Learning and Convolutional Neural Networks

First author · 2022 IEEE Conference on Interdisciplinary Approaches in Technology and Management for Social Innovation (IATMSI), Gwalior, India · in IEEE Xplore, May 2023 · cited 25 times

Trained a CNN to 98.4% accuracy for medical-waste classification and deployed it as an interactive TensorFlow.js web application. Multi-class classification of medical waste as part of an automated system to segregate it without human intervention.

Abstract

With the rise of attention to healthcare since the start of the century, which the recent pandemic has emphasized, the number of hospitals and clinics has increased exponentially. The growth in hospitals and patients has also resulted in increased medical waste. The different kinds of medical waste must be segregated and disposed of properly to prevent the spread of bacteria and viruses and cross-contamination. However, it is not economically feasible to hire a workforce that can segregate said waste. With the perceived popularity of deep learning and image classification systems, creating a Deep Learning model to categorize the different kinds of medical waste is possible. Hence using a deep learning-based classification method in which an appropriate pre-trained model is selected for practical implementation, followed by transfer learning methods to improve classification results, is appropriate. Different types of medical waste are grouped into umbrella categories (general, hazardous, infectious). Three out of four of the modified pre-trained models with different architectures were able to achieve an accuracy above 95 percent.

M. Verma, A. Kumar and S. Kumar, "Medical Waste Classification using Deep Learning and Convolutional Neural Networks," 2022 IEEE Conference on Interdisciplinary Approaches in Technology and Management for Social Innovation (IATMSI), Gwalior, India, 2022, pp. 1-5, doi: 10.1109/IATMSI56455.2022.10119431.keywords: Training; Deep learning; Computer vision; Technological innovation; Hospitals; Computational modeling; Transfer learning; Deep Learning; Medical Waste Classification; Transfer Learning; Convolutional Neural Network (CNN)

IEEE Xplore → PDF · 2.3 MB Code

§03

Projects

ls ~/src

swallow

Swallow

FastAPI
React 19
PostgreSQL

Stop choosing, start watching — rolls a random unseen episode from the shows you follow and never repeats one until the pool is exhausted. Saved “controlled random” presets narrow the roll to specific shows, seasons, and a maximum episode length. FastAPI + SQLAlchemy 2 over Postgres, Clerk auth, React/Vite frontend, episode data from the TVmaze API.

Live Demo →Code

~/src/rumpy — cargo bench

rumpy

Rust
Systems
Benchmarking

A zero-dependency NumPy, written from scratch in pure Rust — same internals as the real thing (flat buffer + shape + strides). Reimplements broadcasting, axis reductions, and matmul from first principles, then benchmarks a cache-friendly ikj loop order against BLAS-backed NumPy. Spec-driven with ~30 milestone tests.

Code

MSM.EXE

msm — YouTube Music CLI

Python
TUI
mpv / yt-dlp

A terminal player that browses YouTube Music and your local library side by side, plays audio through mpv over its JSON IPC socket, and scrobbles to Last.fm via cmusfm. Vim-style keys, live search, play history, and pixelated album art rendered straight in the terminal.

Code

ml-from-scratch

ML-from-scratch

PyTorch
Deep Learning

Core ML/DL models — backpropagation, CNNs, transformers — implemented from first principles in pure PyTorch, without high-level abstractions.

Code

resume-check

Resume Check

NLI
Transformers.js
Client-side

Scores a resume against a hand-written hiring rubric, then matches it against a pasted job description with an open-weights NLI model — entirely in the browser, no upload, no backend.

Live Demo →Code

WASTE.CNN