From messy data to models that work in production.
I'm Aman Tripathi, an ML engineer. I've shipped a model to a production wearable app, released an open-source Python package and published at IEEE, across sensor signals, text, molecules and LLM products. Drive through to see each project.
Sleep detection at Temple
Deep Learning Engineer Intern at Temple (Eternal). When did you fall asleep, and when did you wake? Answered from the PPG and accelerometer of a wearable worn on the temple.
- Shipped a brief-movement detector to the production app
- CNN-LSTM benchmarked against XGBoost, LSTM and Transformer models
- Scored on epoch F1 and onset/offset boundary error
The Faint Signal
An interactive lesson on EEG and brain-computer interfaces, built on a real brain recording. You measure the signal, close the eyes to watch the alpha rhythm bloom, blink to bury it, clean it up, then spell with a BCI. Click the screen beside the road to blink.
- Real EEG from PhysioNet, with filters and ICA shown step by step
- 8 blink-removal methods scored side by side in the Cleaning Lab
- Cerebral Cortex, an AI guide that answers from the site's own pages and cites each one
GlucoTrace
A small Transformer that turns a day of continuous glucose data into a 128-number fingerprint, so similar days can be found and compared. The curve beside the road is a real day from the BIG IDEAs dataset.
- 0.965 similarity kept with 30% of readings missing
- Pre-registered evaluation, splits pinned by checksum
- pip install glucotrace · DOI · 109 tests
Mobile EEG signal quality
Which brain-signal sensors can you trust while someone walks or jogs? A per-sensor score built only from what a wearable cap would have. The head shows my real results at each treadmill speed.
- 10 / 10 people where the score tracked decoding
- 4 of 5 on people sealed away while tuning
- 95% → 0.5% false hum flags after the fix
Wrist PPG motion cleaning
A watch's own accelerometer and gyroscope cancel motion out of the pulse signal. The watch face follows each step below, on a real running recording.
- 2.4× more heartbeats found while cycling
- 0.432 → 0.610 mean beat F1 across activities
- 0.34 s per 5-minute recording, no training
Chat-to-Pay
Small sellers in India take orders as Hinglish WhatsApp messages. Claude reads the order; the app prices it, checks it and sends a payment link.
- The app, not the model, computes every rupee
- Chats are data, so "make it free" can't change an order
- Next.js 16 · TypeScript · Claude API
Drug interactions with GNNs
RDKit turns drug molecules into graphs; a graph autoencoder in PyTorch Geometric predicts whether two drugs interact, including pairs it has never seen.
- Graph autoencoder with link prediction
- DrugBank molecules, ROC-AUC and average precision
- Live app on Streamlit
River-flow forecasting
Variational Mode Decomposition splits a messy, non-stationary series into simpler waves; then 8 models are benchmarked to forecast it, with SHAP to explain the winner.
- 8 models benchmarked, from ARIMA to CNN-Transformer
- NSE 0.985 for the best model, XGBoost
- 50+ features and Diebold–Mariano tests
Benchmarking ML in under-resourced contexts
A peer-reviewed paper on reading public opinion about government policy in South Karnataka from Reddit, when there is very little labelled data. Published at IEEE CIEES 2025 (Ruse, Bulgaria), with Varnica Sharma and Kavitha Mahesh.
- 1,196 Reddit comments analysed
- 5 classical models vs. DistilBERT for sentiment, plus K-Means and BERTopic for themes
- 94.6% K-Means cluster accuracy; the first benchmark for this regional setting
Open-source contributions
I contribute fixes and tests to tools other researchers rely on, and publish my own work so anyone can install it and check it.
- NeuroKit2 · 4 merged pull requests: a respiration fix so
rsp_rvt()finds every breath at any sampling rate (28 of 28 at 100 Hz, up from 10), an HRV tuple-input fix, RQA regression tests and RQA docs - GlucoTrace · on PyPI with a Zenodo DOI, 109 tests and CI on every push
- Public code for mobile EEG quality, wrist PPG cleaning and Chat-to-Pay
Let's work together.
Open to ML engineer, applied scientist, LLM and evaluation roles, remote or anywhere. M.Sc. Data Science, Manipal · B.Sc. CS, Statistics & Maths, Chandigarh University.
amantripathi3273@gmail.com