Skip to content

Chetank190/ProjectHaven

BranchesTags

Repository files navigation

Haven Matrix

NVIDIA Spark Hack Toronto · May 29–31 · Public Services Track

The Spark Story

"Haven Matrix holds all seven data matrices — shelter occupancy, clinical services, food programs, hygiene stations, grassroots resources, public amenities, and 9,255 TTC stops — plus the full Nemotron-30B language model context, simultaneously in the DGX Spark's 128 gigabytes of unified memory.

On any other hardware, you'd be serializing data through a PCIe bottleneck between CPU RAM and GPU VRAM. Here, everything lives in one coherent memory space. When someone speaks to the kiosk, their words become a boolean payload in milliseconds, and cuML solves the nearest accessible resources across all five pillars simultaneously in under ten milliseconds — accounting for whether they have ID, whether they're currently using, whether they have family with them.

That number on the benchmark panel — that's not a claim. That's a measurement from this hardware, right now."

What It Does

Haven Matrix is a dual-gateway triage system for social services in Toronto:

  • Gateway A (Caseworker) — Type or speak client notes → NIM compiles to JSON → constraint-aware GPU KNN → per-client itinerary + warm handoff phone scripts
  • Gateway B (Kiosk) — Voice only → eligibility questions via TTS → spoken care route to resources you can actually enter given your ID status, sobriety, and family situation

Quick Start (Local / MacBook)

# 1. Create and activate a virtual environment
python3 -m venv .vhaven
source .vhaven/bin/activate   # Windows: .vhaven\Scripts\activate

# 2. Install Python dependencies
pip install -r backend/requirements.txt

# 3. Verify datasets
python backend/data_ingestion.py --verify --mode cpu

# 4. Start FastAPI (port 8000)
uvicorn backend.main:app --host 0.0.0.0 --port 8000 --reload

# 5. In a new terminal — start React (port 3001)
cd frontend && npm install && npm run dev

# 6. Open browser
# Caseworker: http://localhost:3001/caseworker
# Kiosk:      http://localhost:3001/kiosk  (requires Chrome for voice)
# Swagger:    http://localhost:8000/docs

NIM/LLM is optional for local dev. Without a GX10 connection, the system falls back to regex keyword matching automatically.

Mac + GX10 setup (team default): backend and frontend on Mac; models on GX10 GPU only → learning/gx10_access_and_gpu_guide.md

Development Architecture (Mac + GX10)

Default: run backend + frontend on your Mac; run models only on the GX10 GPU over Tailscale.

Mac (:3001 frontend, :8000 backend CPU)
    │  NIM_ENDPOINT=http://100.81.85.39:8001/v1
    ▼  Tailscale
GX10 — docker compose up nim -d  (Gemma 3n on GPU :8001)
       optional: docker compose up asr -d  (Parakeet on GPU :9000)
Step Where Command
1. Load model on GPU GX10 (SSH) docker compose up nim -d
2. Backend Mac uvicorn backend.main:app --port 8000 --reload
3. Frontend Mac cd frontend && npm run dev

Mac .env (copy from .env.example):

GX10_TAILSCALE_IP=100.81.85.39          # your unit's Tailscale IP
FORCE_CPU_SOLVER=1
NIM_ENDPOINT=http://100.81.85.39:8001/v1
NIM_FALLBACK=http://100.81.85.39:8001/v1

Model storage: weights download automatically into Docker volume nim-cache on the GX10 on first NIM start (requires NGC_API_KEY in GX10 .env). Nothing is stored on the Mac.

Verify: compile_method: "nim" on a route request; rapids_mode: "cpu" on /api/v1/health.

Full walkthrough: learning/gx10_access_and_gpu_guide.md

Connecting to the GX10 (No Wi-Fi — SSH Required)

The GX10 has no built-in Wi-Fi. Remote access info is on the pamphlet in the box. Connect via mobile hotspot first, then optionally Tailscale for persistent access.

Full guide: learning/gx10_access_and_gpu_guide.md

Step 1 — SSH over mobile hotspot

Connect your laptop to the GX10 hotspot (the GX10 auto-connects to saved hotspot profiles):

Your unit (gx10-3cd8):

Field Value
Hotspot SSID gx10-3cd8
Hotspot Password gx10-3cd8

Open Terminal (Mac/Linux) or PowerShell as Administrator (Windows):

ssh asus@gx10-3cd8.local

When prompted:

  • Type yes and press Enter
  • Password: password

No pamphlet? Flip the unit over → read the MAC1 sticker → use the last 4 characters (e.g. 3C:D8 → gx10-3cd8). Other units use the same pattern: ssh asus@gx10-XXXX.local.

Step 2 — Set up Tailscale (for persistent access across any network)

Install Tailscale on your laptop first: https://tailscale.com/download

  • Do NOT use a .edu email — it blocks registration
  • Mac: allow all prompts, enable from the taskbar icon
  • Windows: enable from the hidden icon tray (right-click)

Then on the GX10 terminal (Tailscale is pre-installed):

sudo tailscale up
# Copy the URL it prints → open in your laptop browser → authorize

After pairing, SSH via Tailscale from anywhere:

ssh asus@gx10-3cd8          # by hostname
ssh asus@100.X.X.X          # by Tailscale IP (shown in Tailscale app)

Invite teammates:

Step 3 — Venue Wi-Fi + remove hotspot (requires monitor)

The hotspot profile persists across reboots. To switch to venue Wi-Fi, connect a monitor to the GX10, pair venue Wi-Fi on the unit, then delete the hotspot profile:

nmcli con show                        # list all connections
nmcli con delete gx10-3cd8-Hotspot   # delete so it doesn't reconnect on reboot

Use the UI from your Mac (default — no port-forward needed)

Backend and frontend run locally on the Mac. Only ensure Tailscale is Connected and GX10 NIM is up. Open http://localhost:3001/caseworker.

If :8001 is unreachable over Tailscale, tunnel the model port:

ssh -L 8001:localhost:8001 asus@100.81.85.39
# then set NIM_ENDPOINT=http://localhost:8001/v1 in Mac .env

GX10 — Models on GPU Only

Backend and frontend run on your Mac. The GX10 only hosts NIM model containers on GPU.

Full guide: learning/gx10_access_and_gpu_guide.md

On the GX10 (SSH)

cd ~/ProjectHaven
cp .env.example .env          # set NGC_API_KEY from build.nvidia.com
docker compose up nim -d        # LLM only — Gemma 3n on GPU :8001
docker compose logs -f nim      # wait until ready (first run: long download)

# optional ASR on GPU
docker compose up asr -d

Verify on GX10:

curl -s http://localhost:8001/v1/models | head
nvidia-smi

Do not run docker compose up (full stack) unless you also want backend on the box.

On your Mac

cp .env.example .env
# set GX10_TAILSCALE_IP and NIM_ENDPOINT=http://<ip>:8001/v1

source .vhaven/bin/activate
set -a && source .env && set +a
uvicorn backend.main:app --host 0.0.0.0 --port 8000 --reload

cd frontend && npm run dev

Verify on Mac: curl localhost:8000/api/v1/health → "rapids_mode":"cpu"; route → "compile_method":"nim".

Optional — Nemotron or GPU KNN on GX10

See gx10 guide appendix and README legacy options only if benchmarking on-box.

Architecture

Mac — React :3001/caseworker | :3001/kiosk
         │ HTTP POST via Vite proxy → localhost:8000
         ▼
Mac — FastAPI :8000 (CPU, FORCE_CPU_SOLVER=1)
         │ HTTP over Tailscale → GX10 :8001
         ▼
GX10 — NIM Gemma 3n :8001 (GPU)  [optional ASR :9000]
         │
         └── Solver on Mac CPU (pandas/sklearn, ~400ms)
              └── 7 datasets + constraint masking

API Endpoints

Method Endpoint Description
POST /api/v1/caseworker/route Text → payload → itinerary + handoff scripts
POST /api/v1/kiosk/session Voice transcript → session + eligibility questions
POST /api/v1/kiosk/route Session + eligibility answers → spoken itinerary
POST /api/v1/caseworker/briefing Morning shelter capacity summary
POST /api/v1/handoff-script Generate caseworker phone call script
GET /api/v1/health Dataset row counts + rapids_mode
GET /api/v1/benchmark Last GPU ms, CPU ms, speedup ratio

Fallback Chain

Component Primary Fallback
LLM Cloud NIM Gemma 3n (if NGC_API_KEY) → Nemotron (llama.cpp :30000) Gemma 3n E4B (NIM :8001) → regex
Data engine cuDF + cuML (GPU) pandas + scikit-learn (CPU — GX10 default)
Shelter data Toronto CKAN (live) Cached data/shelters.csv

Environment Variables

# Mac .env — backend + frontend on Mac, models on GX10 (team default)
GX10_TAILSCALE_IP=100.81.85.39
FORCE_CPU_SOLVER=1
NIM_ENDPOINT=http://100.81.85.39:8001/v1
NIM_FALLBACK=http://100.81.85.39:8001/v1
ASR_NIM_URL=http://100.81.85.39:9000
NGC_API_KEY=your_ngc_key_here          # optional cloud LLM/ASR fallback
NIM_API_KEY=not-needed
VITE_KIOSK_HUB=Union Station

# GX10 .env — on the box only (for docker compose up nim)
# NGC_API_KEY=your_ngc_key_here

Composite Scoring

score = 0.60 × dist_norm + 0.30 × occupancy_ratio + 0.10 × (0 if transit else 1)

Lower score = better match. Balanced between proximity, availability, and transit access.

About

No description, website, or topics provided.

Resources

Readme
Activity

Stars

3 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

 
 
 

Contributors

Languages