arrow_back HD Research Hub
Architecture

How It Works

Four pipelines power the HD Research Hub. All open source, all running on consumer hardware.

PubMed API Google Gemma 4 Google AI Studio Sarvam AI Ollama on Jetson
database

Daily Data Pipeline

Papers flow in from PubMed, get analyzed by LLM on Jetson, and populate the knowledge base that powers everything.

schedule Daily at 8am UTC
api

PubMed API

12 search queries

download

Full Text (PMC)

Open access papers

memory

LLM Analysis

Gemma 4 on Jetson

storage

Knowledge Base

1,571 chunks

public

Live Website

Auto-deploy via Vercel

chat

Ask HD Research (Chatbot)

Questions in any language get translated, matched against full-text research, answered by Gemma 4 via Google AI Studio, and translated back.

verified_user Medical Guardrails

Input

User Question

Any language

Detect

Sarvam AI

22 Indian languages

Search

Knowledge Base

1,571 full-text chunks

Inference

Google AI Studio

Gemma 4 (agentic + multimodal)

Translate

Response Back

User's language

Output

Cited Answer

With PubMed IDs

biotech

Experiment Pipeline

Hypothesis generation and drug repurposing scoring. Runs on the Jetson overnight.

hardware Jetson AGX Orin 64GB
psychology

Question

article

Pull Papers

smart_toy

LLM Analyze

science

Extract

lightbulb

Hypothesize

speed

Score

description

Publish

auto_mode

Autonomous Agents

Four agents run on schedule, keeping the knowledge base fresh and hypotheses updated.

search

Paper Scout

Discovers new papers

Every 6h
auto_awesome

Hypothesis Refiner

Re-scores drug candidates

Daily
visibility

Social Watcher

Monitors X, Reddit, HN

Daily
edit_note

Digest Writer

Weekly research summary

Weekly
menu_book

Wiki Compiler

Compiles knowledge

After each run
rocket_launch

Auto-Publish

Git push triggers Vercel

Instant
all_inclusive

The Self-Improving Loop

Experiments generate data. A wiki compiler organizes the findings. Those findings feed into the next experiment. Each cycle, the system learns from its own output and improves, without anyone having to restart from scratch.

science

Run Experiment

storage

data/*.json

menu_book

Wiki Compiler

auto_stories

wiki/ pages

science

Next Experiment

Every cycle, the wiki grows. 117 targets, 5 hypotheses, 4 experiments compiled so far.

Built With

Google Gemma 4 Google AI Studio Google Cloud Run Ollama Jetson AGX Orin Upstash Vector Sarvam AI Vercel PubMed GitHub Actions

Built by data scientists, not doctors. Nothing on this site is medical advice.