Open to opportunities

Satya Akhil Galla

I architect AI systems — from research to production.

Shipped a productivity plugin with 200 active daily users and a medical AI inference engine running at 0.529s TTFT. IEEE-published in graph neural networks.

↓ scroll
01

Work

Morning OS

Productivity plugin with 200 active daily users — generates a personalized AI briefing each morning from your notes, tasks, and goals.

TypeScriptObsidian APILLM APIs
Obsidian Plugin →
  • 200 active daily users; iterated on features through a live Discord community
  • Multi-provider LLM support with graceful degradation and full mobile compatibility
200 active daily users

Deep Research Agent

Autonomous agent that searches, reads, cross-references, and synthesizes any topic end-to-end without human input — shipped in 5 days.

PythonClaude APIMCP
  • 51 tools across 10 namespaces with subagent orchestration and an LLM-based eval harness
  • Composable tool chains validated across 20+ tool call sequences; shipped from scratch in 5 days
51 tools · 10 namespaces · 5 days

Serverless Clinical AWQ Llama Engine (SCALE)

Medical AI assistant that answers clinical questions from 10,000 documents in under 0.5 seconds — deployed on serverless A10G GPUs.

Llama-3.1ModalvLLMPineconeHuggingFace
Try it live →
  • 0.529s TTFT and ~60 tok/s on A10G — fast enough for real clinical use
  • 14% USMLE accuracy gain from AWQ fused kernels over NF4 baseline
0.529s TTFT · ~60 tok/s · 14% accuracy gain

Lunar GCN — ISRO / IEEE WHISPERS 2023

Graph Convolutional Network that classified lunar surface types from Chandrayaan hyperspectral data, beating CNN baselines by 13.5%.

GCNPyTorchHyperspectral ImagingNASA Chandrayaan
Read paper →
  • +13.5% accuracy over CNN baseline on Chandrayaan-1; 91% accuracy on Chandrayaan-2 data
  • Published at IEEE WHISPERS 2023 and presented as a NASA NESF 2023 poster
+13.5% vs CNN · 91% accuracy · IEEE published

CoCo: Neuro-Symbolic Desktop Companion

Desktop AI agent that watches your activity in real time and adapts to your emotional state — fast reactive layer and slow reasoning running in parallel.

Phi-3PyTorchTCP/IPCAgentic AI
  • <16ms UI latency at 60Hz — imperceptibly fast reactive loop
  • Runs fully on-device; zero data leaves the machine
<16ms UI latency · 60Hz reactive loop
02

Experience

Graduate Teaching Assistant

Boston University

CS 365: Foundations of Data Science

Sep 2025 – Jan 2026
  • Led technical labs on Python optimization and statistical modeling, mentoring 50+ students in data science fundamentals including Pandas, NumPy, and algorithmic best practices.
  • Facilitated weekly code reviews to enforce production-standard coding styles and debugging techniques.

Machine Learning Engineer

Aaizel International Tech

Internship

Feb 2024 – May 2024
  • Architected a multimodal fusion pipeline (IFCNN) merging IR+RGB satellite imagery and constructed the domain's first remote sensing scene graph dataset.
  • Fine-tuned a RelTR Transformer for geospatial relationship extraction and deployed the inference engine to Microsoft Azure, enabling real-time semantic visualization.

Computer Vision Engineer

Terrafic Inc

Internship

Jun 2023 – Nov 2023
  • Integrated the Segment Anything Model (SAM) and OpenCV to automate maritime feature extraction; delivered a production-ready MVP that secured initial client pilots.
  • Implemented Super-Resolution (SR) algorithms (4x/6x) to enhance object detection performance on low-resolution defense mapping data.
03

Skills

Daily use

PythonTypeScriptLangGraphClaude APIvLLMRAG ArchitecturesDockerGitCursorClaude Code

Projects & research

Llama-3Phi-3 (Quantization)AWQ / NF4ModalPineconeAWS (EC2/S3)AzureFastAPIGCNsPyTorchOpenCV
04

Background

Efficient Graph Formulation and Latent Space Integration for Lunar Hyperspectral Image Classification

2023

IEEE WHISPERS 2023

GCN-based pipeline for Chandrayaan-1 hyperspectral data, achieving 90.1% accuracy on lunar surface classification via adaptive graph construction and spectral-spatial fusion.

Enhancing Hyperspectral Classification through GCNs with Adaptive Graph Construction

2023

NASA NESF 2023 (Poster)

Adaptive graph construction method for Chandrayaan-2 IIRS hyperspectral data, reducing false positive rates by 13.5% through improved spectral neighbor modeling.

Education

M.S. in Artificial Intelligence

Boston University

GPA: 3.7 / 4.0

Jan 2026

B.Tech. in Computer Science

IIIT Sri City

GPA: 3.5 / 4.0

May 2024
05

Contact

Let's build something.

Open to full-time roles, research collaborations, and interesting projects.

or press ⌘K and type 'email'