Sugeerth Murugesan
Staff AI Scientist with 8+ years building LLM systems and production machine-learning pipelines — driving a 31% engagement lift, 17% revenue growth, and $4M+ cost savings across products serving 5M+ users. Career arc spans the modern ML stack: foundational CNNs for video recognition (Samba TV), interpretable deep learning (Adobe Research), and transformer fine-tuning, evaluation and deployment on distributed systems at scale (Intel / Intuit). PhD in Computer Science, UC Davis.
Experience
- Fine-tuned transformer models powering LLM-based recommendation systems in production, leading to a 31% engagement lift and 17% revenue growth across 5M+ users.
- Built and operated agentic AI systems end to end — retrieval, tool use, evaluation harnesses and human-escalation policies — on distributed serving infrastructure (vLLM / SGLang).
- Led LLM post-training programs (DPO, GRPO, RLHF, LoRA/QLoRA) with eval-gated releases: golden sets, trace-level metrics and regression gates in CI.
- Designed DeepInsight, a real-time ML analytics platform (Kafka, Flink, Kubernetes; C++/Go/Python) processing production event streams for anomaly detection, contributing $4M+ in cost savings.
- Scaled training with distributed machine-learning pipelines (DDP, FSDP, DeepSpeed ZeRO), profiling quality/cost Pareto frontiers to cut serving spend without metric loss.
- Co-created Sambavision: shipped applied computer-vision research into product, constructing foundational CNNs for large-scale video recognition on smart-TV content streams.
- Built DeepCompare, an interpretable-deep-learning system for comparing model behavior — resulted in a patent and an IEEE CG&A publication.
- Computational Research Division: visual analytics and ML over large-scale scientific data; four peer-reviewed journal publications (IEEE TCBB, BMC Bioinformatics, ACM-BCB).
Selected Projects — all open-source, live at sugeerth.github.io
- Prompt Studio — zero-cost, client-side prompt agents: a live dual-encoder intent model scoring 20 intents per keystroke, adaptive multi-variant prompt composition, and optional on-device WebLLM inference.
- GPU Training Lab — ~25 runnable notebooks covering the modern post-training stack end to end: DPO, GRPO reasoning, RLHF, LoRA/QLoRA and distributed DeepSpeed training.
- Self-Learning LLM Training — a Trainer → Judge → MetaJudge hierarchy that drives, evaluates and critiques its own training runs, with bias audits and human escalation on close calls.
- Agent evaluation research — published interactive essays on trace-level evaluation of long-horizon agents (recall@K, progress-AUC, behavioral clustering), agent memory and context management, inference-time memory bandwidth, autonomy policy and tool design.
Skills
AI / ML: Transformer models, LLM post-training (DPO · GRPO · RLHF), LoRA/QLoRA fine-tuning, RAG, recommendation systems, multimodal VLMs, graph neural networks, PyTorch, TensorFlow
Evaluation: LLM-as-judge, eval harnesses & golden sets, trace-level agent metrics, experiment design & ablations, quality/cost Pareto analysis
Systems: Distributed systems & training (DeepSpeed · FSDP · DDP), inference serving (vLLM · SGLang), machine-learning pipelines (Kafka · Flink), Kubernetes, AWS, agent observability & tracing
Programming: Python, C++, Go, TypeScript/JavaScript, SQL
Education
- PhD, Computer Science — University of California, Davis · Research Associate, UC Berkeley / Lawrence Berkeley National Laboratory
- B.Tech, Computer Science — Amrita University (with UC Davis)
Publications & Patents
- Brain Modulyzer — multi-view analysis of functional brain networks, IEEE TCBB
- ECoG ClusterFlow — hierarchical clustering of dynamic brain networks, BMC Bioinformatics / ACM-BCB
- DeepCompare — interpretable comparison of deep models, IEEE CG&A + US patent
- Ongoing technical writing: interactive essays on agent evaluation, memory and infrastructure — sugeerth.github.io
This page is the canonical, always-current resume. Use Print / save as PDF above for a clean letter-size copy — it prints without the navigation or buttons.