Graduate Research Assistant

NASA IMPACT August 2024 - Present
Current Role

Contributing to NASA's Accelerated Knowledge Discovery project, building the model-serving infrastructure and ML pipelines behind a deep research LLM agent with human-in-the-loop guidance for scientific exploration.

Key Achievements:

  • Owned end-to-end design and deployment of a scalable multi-model inference service in Python and FastAPI, serving multiple ML models through a single endpoint with dynamic resource management; Redis-based distributed task queuing reduced API response latency by 40%
  • Designed a two-level guardrails system (IBM Granite Guardian served via local Ollama inference as a first-pass filter, followed by a dynamic Risk Agent), achieving ~80% cost reduction over single-model approaches
  • Fine-tuned and evaluated classifiers (Indus, ModernBERT, BERT, FastText) on large scientific corpora, reaching 98% precision on multi-label division tagging and 96% recall on single-label indexing, with conformal prediction (α = 0.05) confidence filtering to reduce false negatives
  • Architected multi-agent workflows for NASA's Deep Research Agent (LLM orchestration, multi-rubric evaluation, and RAG-based agents for code, data, and gap search) and built MCP tools and servers with FastMCP exposing NASA's Science Discovery Engine
  • Designed benchmark evaluations for agent architectures using synthetic datasets and A/B testing across configurations (tools, web search, refined prompts), and benchmarked IBM Granite Guardian models against dynamically generated risk criteria
  • Built production data pipelines with custom web resolvers and scrapers, improving data acquisition throughput by 3x
Python FastAPI Model Serving Distributed Systems Redis Ollama LLM Agents MCP RAG PyTorch
NASA IMPACT

Machine Learning Engineer

MutualArt September 2023 - June 2024

Built GPU-backed inference services and ML-powered features for an art market intelligence platform serving thousands of users.

Key Achievements:

  • Fine-tuned IDEFICS-2, an 8B vision-language model, with LoRA (rank-16, ~20M trainable parameters, 0.25% of total) and 4-bit NF4 quantization via BitsAndBytes with flash attention, reducing VRAM requirements by ~75% while maintaining accuracy
  • Developed a production microservice on GPU servers consuming AWS SQS events for automated description generation and validation with BLIP Image-Text Matching, improving content discoverability by 80% while maintaining 99.5% uptime
  • Engineered a multimodal embedding pipeline fusing text, categorical, and numerical features into unified 424-dim vectors; deployed as containerized services on AWS ECS/Fargate processing 80,000+ artworks weekly, with Qdrant ANN retrieval and OpenSearch scoring combined via Reciprocal Rank Fusion for real-time recommendations
  • Built a parallel image-processing pipeline for artwork duplicate detection using two-stage ORB + SSIM matching with Union-Find grouping, improving customer conversion by 200%
  • Separated the lightweight recommendation service (Qdrant and OpenSearch lookups) from the compute-heavy embedding service (Sentence Transformer inference) for independent scaling and cost efficiency
PyTorch LoRA Quantization GPU Inference AWS ECS/SQS Qdrant Computer Vision
MutualArt

Associate Software Engineer

Sireto Technology September 2022 - September 2023

Developed KuberIDE, a web-based IDE for Cardano developers within a microservices architecture. Led backend development focusing on authentication, access management, and deployment optimization.

Key Achievements:

  • Led development of KuberIDE, a web-based IDE (ReactJS, FastAPI) within a microservices architecture serving 500+ developers, reducing environment setup time from hours to minutes
  • Designed and implemented backend authentication, access management, and session handling using OAuth 2.0 cookies for browser-based access and API keys for secure multi-tenant direct access
  • Built an HTTPS gateway to in-house microservices and a WebSocket proxy providing real-time Haskell IntelliSense via the Haskell Language Server, increasing platform adoption by 300%
  • Optimized deployment with Docker multi-stage builds (40% image size reduction), Pytest automated testing (80% coverage), and GitHub Actions CI/CD pipelines
  • Mentored 2 junior developers on API design patterns and testing best practices through code reviews and pair programming
Python FastAPI ReactJS Docker CI/CD OAuth WebSocket
Sireto Technology

Solutions Engineer Intern

Logpoint January 2022 - August 2022
Internship

Focused on Linux system administration, performance optimization, and automation. Gained extensive experience in system monitoring, troubleshooting, and scripting for enterprise-level infrastructure.

Key Achievements:

  • Diagnosed and resolved critical incidents on enterprise Linux systems, monitoring and tuning for performance and reliability
  • Automated recurring operations with shell scripting, reducing manual intervention across the infrastructure
Linux Shell Scripting System Administration Performance Optimization
Logpoint

AI Research Intern

NAAMII May 2020 - August 2020
Research

Research internship focused on Computer Vision and AI model development. Worked with state-of-the-art CV models and contributed to various projects involving object detection, facial recognition, and image segmentation.

Key Achievements:

  • Built and fine-tuned computer vision models for object detection, facial recognition, and image segmentation using PyTorch, OpenCV, and scikit-learn
  • Collaborated with lead researchers to implement and improve state-of-the-art CV models across active research projects
Python PyTorch OpenCV Computer Vision scikit-learn
NAAMII