SEEKING WORK | Bengaluru, India (UTC+5:30) | Remote only | Contract / fractional, ~20 hrs/week
Senior ML engineer, 10 years. The last 6 owning the entire model layer of a mobile DSP - win-rate prediction, pCTR/pCVR, install prediction, predictive audiences, bid pricing. Models served in-process inside a live bidder under a ~200 microsecond latency ceiling, across ~1B daily bid requests for ~200 advertisers.
Things I shipped there:
- Replaced a parametric win-rate curve-fit with a discriminative neural classifier predicting P(win | inventory features, bid price). The bidder uses it to forecast spend and halt bidding against budgets on a 10-minute refresh.
- Per-advertiser XGBoost conversion models (BigQuery ML) replacing hand-written rule segments. Top-decile audiences converted at 15-20%, AUC >0.90, scaled to 60-70 models in production. It moved the client mix - gaming went from ~5-10% to ~40-50% of the book.
- CPM cap model that inverts the win-rate curve to price for a target win rate by geo/OS/format. ~10% margin improvement.
- Calibrated bidding chain, pCTR x pCTC -> pCVR with isotonic regression, because direct bid pricing only works if the probabilities are honest.
- Install prediction for iOS Limit Ad Tracking users - modelling conversion with no persistent identifier.
- Earlier: backend services and big-data pipelines at 4TB/day (Hadoop, Hive, Spark).
Good fit for: ranking and recsys, CTR/CVR, auction or marketplace pricing, experimentation design where clean randomisation isn't possible, and getting a model out of a notebook into production under a real latency budget.
I'm building my own market-data platform right now (~8,300 Indian equities, point-in-time correct, running unattended on GCP), so this is deliberate part-time capacity.
Things I shipped there:
- Replaced a parametric win-rate curve-fit with a discriminative neural classifier predicting P(win | inventory features, bid price). The bidder uses it to forecast spend and halt bidding against budgets on a 10-minute refresh.
- Per-advertiser XGBoost conversion models (BigQuery ML) replacing hand-written rule segments. Top-decile audiences converted at 15-20%, AUC >0.90, scaled to 60-70 models in production. It moved the client mix - gaming went from ~5-10% to ~40-50% of the book.
- CPM cap model that inverts the win-rate curve to price for a target win rate by geo/OS/format. ~10% margin improvement.
- Calibrated bidding chain, pCTR x pCTC -> pCVR with isotonic regression, because direct bid pricing only works if the probabilities are honest.
- Install prediction for iOS Limit Ad Tracking users - modelling conversion with no persistent identifier.
- Earlier: backend services and big-data pipelines at 4TB/day (Hadoop, Hive, Spark).
Technologies: Python, SQL, TensorFlow, XGBoost, BigQuery/BQML, Vertex AI, GCP, Spark, Hive, Hadoop, scikit-learn, Keras.
Good fit for: ranking and recsys, CTR/CVR, auction or marketplace pricing, experimentation design where clean randomisation isn't possible, and getting a model out of a notebook into production under a real latency budget.
I'm building my own market-data platform right now (~8,300 Indian equities, point-in-time correct, running unattended on GCP), so this is deliberate part-time capacity.
Resume/CV: https://drive.google.com/file/d/1lpBJK5koZuWOwjPfRMWW4nDod8f...
Email: garg17793@gmail.com
LinkedIn: https://www.linkedin.com/in/parshant-garg