Initializing particles
Wiring synapses
Open to ML & AI roles · St. Louis, MO
01 — About
I'm an AI/ML engineer finishing my M.S. in Artificial Intelligence at Saint Louis University. I build machine learning systems end to end — data pipelines and feature engineering, model training in PyTorch, containerized inference APIs, and cloud deployment with CI/CD and monitoring.
Right now I evaluate frontier LLMs and coding agents, build a vision-language model benchmark as a research assistant at SLU, and lead a capstone on uncertainty-gated model routing for LLM tool calling. Before that: deep learning on satellite imagery and an explainable clinical AI platform.
GPA · M.S. AI
Years applied ML & stats
Certifications
Awards & scholarships
02 — Experience
Each role added a layer: evaluating frontier models, researching new ones, and shipping systems that hold up in production.
AI Trainer — LLM & Coding-Agent Evaluation · Contract · Remote
Research Assistant — Multimodal (VLM) Evaluation · St. Louis, MO
Graduate Researcher — Deep Learning & Computer Vision · St. Louis, MO
Applied Statistics Specialist · Remote
Research Intern — IAS Summer Research Fellow · Mumbai, India
03 — Selected work
Keep scrolling: the track moves sideways,
one project at a time.
Sep 2026 — M.S. Capstone
Uncertainty-gated small-model routing for LLM tool calling
Serves tool-calling requests with small open-weight models and escalates only the uncertain calls to a 14B fallback. Confidence comes from token log-probabilities and multi-sample tool-call disagreement. The goal is 20% lower cost per successful task while keeping 95% of the 14B model's success rate.
Python · PyTorch · vLLM · Qwen2.5 / Llama-3.2 · LangGraph · MCP · FastAPI
View on GitHubSpring 2026
Explainable AI clinical risk prediction platform
Full-stack cardiovascular risk platform with FastAPI model serving, role-based access and per-patient SHAP attributions. A RAG layer grounds LLM explanations alongside, but strictly separate from, the model's risk score.
FastAPI · React · PostgreSQL · XGBoost · SHAP · AWS
View on GitHubFall 2025
Multi-class object detection with LW-DETR
Collected and annotated a custom 3,000+ image dataset across 8 categories, trained an LW-DETR detector with transfer learning, and built class-wise error-analysis tooling to diagnose failure modes.
PyTorch · LW-DETR · COCO · Augmentation
View on GitHub2025
Traditional dice game rebuilt as an agent testbed
A traditional Andhra Pradesh dice game turned reinforcement-learning playground, moving from rule-based play to a Q-learning opponent trained through self-play.
Java · Q-Learning · Reinforcement Learning
View on GitHub2025
Multi-modal music recommender
Upload a photo and get a playlist. Image captioning plus mood inference turns what the camera sees into music recommendations through the Spotify API.
BLIP · DistilGPT2 · Spotify API · Streamlit
View on GitHub04 — Research
Do vision-language models credit the right source? A controlled photo–caption benchmark with a pre-registered protocol that checks whether VLMs take event actors from the image or the caption, and when they make evidence up.
CNN, GAN and attention-based models for restoring shadowed regions in aerial imagery, and measuring how that restoration changes downstream feature quality.
Multi-scan point-cloud registration with ICP and RANSAC. The terrain behind this text is a nod to that work.
Honors & leadership
Awards
IIT Bombay
Chegg India
Vel Tech
75% tuition waiver (B.Tech)
Leadership
All India University South Zone Tournament
Indian Concrete Institute
National Service Scheme
05 — Skills
Hover a group and its cluster lights up in the scene.Each group maps to one object in the scene.
Certifications