Aswin Jose

Aswin Jose

ML engineer. AI for health, and building things for fun.

I am an engineer in Bengaluru, working remotely for Sonus Health. Most of what I do is audio that was never meant to be analysed: heart sounds picked up by an ordinary phone microphone. I work well beyond that too, across computer vision, graph machine learning, backend development, model optimisation and database systems.

  • Machine learning engineer with around five years across research and engineering, including two years full time in industry, building production ML systems and healthcare AI.
  • AI product engineering, from framing the problem through prototyping and validation to putting it in front of real users.
  • Production ML: model optimisation, inference acceleration, calibration, evaluation frameworks, monitoring and deployment.
  • Generative AI: LLM orchestration, retrieval-augmented generation, embeddings, vector search, and human-in-the-loop workflows.
  • Backend and distributed systems: high-throughput asynchronous services, caching, streaming and database-driven applications at a million daily active users.
  • Performance work: inference acceleration and latency reduction, including roughly a two and a half times speedup over Treelite on tree-ensemble inference.
  • Research-driven: four publications in machine learning for healthcare and cancer genomics, including work in Briefings in Bioinformatics.
  1. Machine Learning Engineer Sonus Health / Decorte Future Industries Mar 2025 to present. London, remote from Bengaluru.

    I own the machine learning and the backend that serves it: the audio-analysis engine behind a live smartphone cardiac screening product, detecting heart murmurs from raw phone-microphone audio, with probability calibration and confidence tiers rather than a bare yes or no, held together by golden test suites and frozen output contracts. I also built the multi-step language-model pipeline that drafts clinical reports, with a clinician reviewing every one before it is signed. A separate pilot with a major European automotive manufacturer looked at pulling vital signs from seatbelt-mounted microphones in a moving car.

  2. Software Engineer Couture.ai Jun 2024 to Mar 2025. Bengaluru.

    Search and personalisation for an e-commerce platform at a million daily active users. I built the retrieval-augmented pipeline end to end, embeddings through vector search to ranking, as an asynchronous high-throughput service under tight latency limits, and deployed it with customer engineering teams.

  3. Machine Learning Intern VoltaML Jul 2022 to Oct 2022. California, remote.

    Inference acceleration. Compiling tree ensembles with the LLVM-based lleaves compiler made XGBoost inference roughly two and a half times faster than Treelite.

  4. Undergraduate Researcher CCNSB, IIIT Hyderabad May 2021 to Jun 2024. Hyderabad.

    Graph machine learning on multi-omics data for patient survival prediction, using graph autoencoders, variational autoencoders and network diffusion. Most of my published work started here.

  5. Software Developer Intern Administration, IIIT Hyderabad Jan 2021 to Apr 2021. Hyderabad.

    Led a team building a full-stack help-query system for the institute, in Node.js, Express, React and PostgreSQL, wired into the existing single sign-on so staff and students did not need another account. I also adapted a wiki engine to hold the institute's internal documentation.

Score to MIDI Photograph sheet music, get something you can play. Python, FastAPI

Photograph sheet music, get a MIDI file and a playable piano roll, with a practice mode that waits for you to hit the right note.

Read more
SimpleRA A relational database engine, extended in C++. C++

A relational database engine extended with new data types including matrices, optimised storage operations, page hashing and locking.

NooooSH A Unix shell written from scratch, no libraries. C

A Unix shell written from scratch in C, using no third-party functions.

Source
BetterSplit A blockchain currency that untangles group debts. Python, Flask

A cryptocurrency on a blockchain, with an algorithm for simplifying the web of transactions inside a group, wrapped in a web app.

gDrug Predicting which drugs bind which targets. Python, PyTorch

Predicting which drugs bind which targets, feeding several similarity matrices into a fully connected network. Adding graph-level structure for both the drug and the target is what moved it past the baselines it was measured against.

Lung segmentation Res-UNet segmentation, built to actually deploy. Python, PyTorch

A Res-UNet for segmenting lung images, benchmarked against alternative pipelines. Most of the effort went into making it deployable rather than only accurate, comparing architectures and backbone configurations for one that stayed usable outside a notebook.

fMRI analyser Reading brain scans as video to infer actions. Python, PyTorch

Relating fMRI scans to the actions a subject was performing, using a video transformer so the scan sequence is read as a time series rather than as unrelated volumes.

Network-based stratification Grouping patients by pathway similarity. Python

A variant of network-based stratification adapted to multi-omics data, propagating each patient's profile over a gene network so that patients group by pathway-level similarity rather than by raw feature overlap. The method behind the renal carcinoma survival paper below.

Source
Trade data infrastructure Market data with point-in-time correctness. Python, PostgreSQL, Docker

Collects and stores OHLCV, news, fundamentals and economic data across global markets, with symbol discovery, historical backfill and point-in-time snapshots so a backtest only ever sees what was knowable on the day.

Trading analytics toolkit Correlation and prediction over 26 years. Python

Correlation analysis and price prediction over roughly 2,300 Indian stocks and 26 years of daily records, looking for leading indicators and cross-market effects between US and Indian equities.

Languages
Python, C++, C, JavaScript, R
Machine learning
PyTorch, XGBoost, ONNX, model calibration, model serving, inference optimisation, latency tuning
Language models
LLM pipelines and orchestration, retrieval-augmented generation, embeddings, vector search, Qdrant, human-in-the-loop systems
Domains
Audio signal processing, healthcare and clinical AI, search and personalisation, multi-omics, graph autoencoders, network diffusion
Backend
FastAPI, Node.js, asynchronous service design, distributed systems, REST APIs, Kafka, Redis, PostgreSQL, MongoDB
Infrastructure
Docker, Kubernetes, AWS, Cloudflare, Git, shell
Practices
Production ML, deployment, monitoring, benchmarking, automated testing, reliability engineering
  • B.Tech in Computer Science, and MS in Computational Natural Sciences

    IIIT Hyderabad, 2019 to 2024. Institute Research Award, 2022 to 2023.

← Back