Sukanya Moorthy

Stealth mode—

Expertise

LLMOps tooling, post-training, data augmentation infrastructure, platform architecture.

Production pipelines, framework migrations, model serving at scale.

Relation extraction, weak supervision, knowledge graphs.

Guardrails, out-of-scope detection, fact verification, red teaming.

LLM guardrails LLMOps fine-tuning knowledge graphs relation extraction weak supervision TensorFlow / Keras Triton red teaming

Work

Sr Machine Learning Engineer II

Mar 2022 — Nov 2025

Credit Karma · Oakland, CA

Founding member of the 4-person team that set Credit Karma's long-term GenAI architecture; the design scaled into a dedicated org and platform strategy.
Unblocked a stalled high-priority initiative with a new systems design: ~$1M/year saved.
Ran the TensorFlow 1.x to 2.x migration across 15+ teams: ~90 production models retrained, preprocessing 30% faster, training time down 20%.
Launched the LLMOps platform and data augmentation infrastructure the GenAI org builds on.
Architected the safety layer for Credit Karma Assistant: out-of-scope detection and fact verification pipelines in the response path.
Cut model launch timelines 50%.
Beyond Credit Karma, volunteered with the Intuit Research team on temporal memory for fintech: built the pipeline extracting temporal events from financial documents and led a comparative analysis of memory frameworks — Zep, Mem0, HippoRAG, Letta.

Mar 2017 — Mar 2022

Yahoo Research · Sunnyvale, CA

Built the relation extraction pipeline turning raw Wikipedia text into subject–predicate–object triples — a transformer model trained with distant supervision, 85% accuracy on the top 15 relations. Knowledge graph coverage up ~73%.
Built the sports event and rumor extraction system, served in production through Triton — bootstrapped training data from a small editorial seed set with weak supervision, shipped a multi-label CNN topic classifier and player-level timelines. CTR up 105%, conversion up 7.74%.
Extended the BRAT annotation tool into a labeling workflow adopted by Yahoo's data teams.

Review & judging

Role	Venue
Peer reviewer	NeurIPS, EMNLP / ACL ARR, ECCV, LogConf
Technical reviewer	O'Reilly (Designing Distributed Systems, 2nd ed., by Brendan Burns, Kubernetes co-founder — AI and inference serving chapters), Manning (Hugging Face in Action)
Top 30 nationwide, in-person finalist	NIST ARIA red-teaming with Humane Intelligence. Contributed to the NIST report and the EvalEval poster at UCSD.

Talks & patents

Education

Off the clock

scroll →