Sampreeth Avvari
Resume

Bay 036 projects · 5 articles

Machine Learning Engineering.

Models are easy; pipelines that keep them honest are the work.

I ship models with the unglamorous parts done properly, like a dental CT checker built in 37 days that cost $62 on Google Cloud in October, against a six figure vendor quote.

Scroll to walk the bay

How I work

  • The model is the small part. A frozen dental segmentation encoder runs on CPU in 2.2 seconds; the rest is the data pipeline, the holdout and the threshold.
  • One transform for train and serve. In Loan Radar, training and the scoring API load the same saved transform, fit on training rows only.
  • A holdout that stays out. A frozen 20 scan holdout stays out of training, and it caught a leaky 0.80 AUROC.
  • Six models, one interface. Six candidates sat behind one interface by April 7 before Model H won.
  • Thresholds priced in rescans. A missed bad scan costs 15 times a false flag, and the threshold is set from that.
  • Depth on the training side. At NYU I trained Llama 3 8B to argue with SFT, a reward model and GRPO.

What I bring

Medical imaging, RLHF research, and MLOps that survives contact with production.

  • No train and serve skewTraining and serving load one saved transform, fit on training rows only.
  • Honest evaluationA holdout kept out of training, and the honest number kept when a leaky one looked better.
  • Training side depthSFT, reward models and GRPO at NYU; a frozen segmentation encoder with a tuned head at Hybridge.
  • Cost aware servingScale to zero Cloud Run, $62 measured from October 1 to 28 against a six figure vendor quote.
In testing

CBCT Scan Validator

Flags bad dental CT scans before a designer opens the case: good or bad, an artifact family, and a class from rules.

  • Six candidate models behind one interface; a frozen DentalSegmentator encoder with a small head won
  • Event driven: the SFTP box hands each scan to the validator, and a real zip went end to end in 26.7 seconds
  • Of 30 studies the testing team marked bad, the review app's model flagged 26
Guided tour
In production

Daily Sales Report

Forecasts lab revenue and new arch starts for this month and the next two, calibrated to about 68% coverage.

  • Current month revenue error falls from 4.2% early in the month to 1.5% late
  • No auto tuning: the best in sample grid lost 0.62 points out of sample, so it stayed out
  • A better lag model waits behind a switch until its gain clears the noise floor
Research

LLM Persuasion (NYU research)

Training Llama 3 8B to argue: SFT, a reward model trained on ChangeMyView deltas, then GRPO against that judge.

  • Preference pairs from ChangeMyView: a reply that earned a delta against one that did not
  • 4 bit Llama 3 8B under every stage, with LoRA rank 64 doing the learning
Academic

Loan Radar

MLOps for a loan default model on Chameleon Cloud: containerized ETL, MLflow on Postgres and MinIO, a FastAPI scorer and stress tests.

  • Training and serving load one saved transform
Academic

ResNet under 5M parameters

CIFAR-10 under a hard parameter cap: augmentation took the model from 93.76% to 97.12% without adding a parameter.

Academic

Movie twins at 33.8M ratings

MinHash and LSH found movie twins among 330,975 users after Spark kept losing nodes, beside a Spark ALS recommender.

The writingAll posts

A dental CT scan checker in 37 days, on CPU, for $62 in OctoberMay 2026 · 5 min read The finance forecast now builds itself by 6:25 every weekdayOct 2026 · 5 min read I taught Llama 3 to argue with a judge trained on Reddit deltasSep 2025 · 4 min read Loan Radar: training and serving that load one saved transformMay 2025 · 4 min read Augmentation took our capped ResNet to 97.12% on CIFAR-10Dec 2024 · 4 min read

End of the bay

Where next.

Get in touch Every project Resume The station