-
-
SURYA KARTHI Ground Truth Edge-OI Platform
💡 Inspiration & Problem Statement
In sub-Saharan Africa, neonatal mortality accounts for over 40% of all under-five deaths. Rural primary health posts and dispensary clinics operate under severe infrastructural deficits: zero reliable grid connectivity, chronic power outages, and an acute shortage of specialized pediatricians.
While cloud-based Foundation Models have demonstrated medical reasoning capabilities, they are completely non-viable at the disconnected edge. Furthermore, purely probabilistic generative models present fatal risks in acute pediatric triage—frequently hallucinating standard adult drug dosages or miscalculating renal clearance in preterm neonates.
SURYA KARTHI was conceived as an offline, air-gapped Ground Truth Edge-OI (Organic Intelligence) Platform—a zero-cloud bedside clinical copilot engineered to operate deterministically on budget 8 GB laptops within a constrained 15W solar power envelope.
⚡ What It Does
SURYA KARTHI delivers real-time clinical stabilization guidance and weight-adjusted pharmacology for vulnerable preterm neonates presenting with respiratory distress, hypothermia, and neonatal sepsis.
- Deterministic Fast-Path Reflex (0.013 ms): Intercepts patient vitals ($SpO_2$, Heart Rate, Respiratory Rate, Birth Weight, Postmenstrual Age) and evaluates acute physiological shock indices and pediatric renal function using closed-form mathematical equations before executing LLM compute.
- Deliberative Slow-Path Synthesizer (Qwen 2.5 3B): Orchestrates end-to-end, non-invasive triage protocols without requiring supplemental oxygen infrastructure (tactile airway stimulation, Kangaroo Mother Care thermal regulation, and hydration-adjusted antimicrobial therapy).
- Zero-Hallucination Pharmacology Core: Anchors antibiotic dosing (Ampicillin, Gentamicin) to empirical Hill-equation GFR maturation curves, preventing aminoglycoside-induced ototoxicity and acute kidney injury.
🏗️ Technical Architecture: Carbon-Silicon Symbiosis
┌──────────────────────────────────────┐
│ Bedside Telemetry & Vitals │
│ (Preterm 850g, SpO2 84%, Grunting) │
└──────────────────┬───────────────────┘
│
┌────────────────────┴────────────────────┐
▼ ▼
┌─────────────────────────────────┐ ┌─────────────────────────────────┐
│ FAST-PATH REFLEX ENGINE │ │ SLOW-PATH DELIBERATIVE CORE │
├─────────────────────────────────┤ ├─────────────────────────────────┤
│ • Latency: 0.013 ms (C++ / Num) │ │ • Model: Qwen 2.5 3B (Q4_K_M) │
│ • Hill-Equation GFR Calculation │ │ • Runtime: llama.cpp (Local CPU)│
│ • Instant Nephrotoxic Overdose │ │ • Structured Clinical Playbooks │
│ Intercept & Red Alert │ │ • Non-Invasive Oxygen-Free Care │
└────────────────┬────────────────┘ └────────────────┬────────────────┘
│ │
└────────────────────┬────────────────────┘
▼
┌──────────────────────────────────────┐
│ Verified Bedside Protocol │
│ (UI Streaming & Telemetry Display) │
└──────────────────────────────────────┘
1. The Fast-Path Reflex Core
- Hill-Equation GFR Clearance: Models developmental renal maturation as a function of postmenstrual age (PMA) and weight:
$$\text{GFR} = \text{GFR}{\text{adult}} \cdot \frac{\text{PMA}^{\gamma}}{\text{PMA}{50}^{\gamma} + \text{PMA}^{\gamma}} \cdot \left(\frac{\text{Weight}}{70}\right)^{0.75}$$
- Eliminates generative hallucinations for life-critical therapeutic windows.
2. The Slow-Path Local Engine
- Model:
Qwen2.5-3B-Instructquantized to GGUF Q4_K_M. - Runtime: Fully optimized
llama.cppbinary bundle running across 4 CPU threads with zero external network socket calls.
📊 Benchmarks & Edge Profiler Telemetry
Validated locally using the official adtc-profiler in participant mode:
| Metric | Measured Value | ADTC Constraint Target | Status |
|---|---|---|---|
| Active Memory Footprint (RSS) | 2.14 GB | $< 8.0\text{ GB}$ (Budget Laptop) | PASS ✅ |
| Fast-Path Reflex Intercept | 0.013 ms | $< 10.0\text{ ms}$ (Sub-second) | PASS ✅ |
| 15W Solar Edge Throughput | 8.12 TPS | $> 5.0\text{ TPS}$ (Solar Throttle) | PASS ✅ |
| Standard Laptop Compute Throughput | 88.74 TPS | N/A (Standard Profile) | PASS ✅ |
| Deliberative TTFT | 1.24 s | $< 5.0\text{ s}$ | PASS ✅ |
| Network Packets During Inference | 0 (Air-Gapped) | Zero External Dependencies | PASS ✅ |
🛠️ How We Built It
- Core Runtime:
llama.cpp+ Python bindings for deterministic edge execution. - SLM Backbone:
Qwen2.5-3B-Instruct(Q4_K_M GGUF). - Clinical Mathematics: Custom NumPy & C-accelerated pharmacokinetics module for Hill-equation GFR and organ maturation modeling.
- Bedside Interface: Streamlit dark-mode clinical dashboard styled with custom high-contrast CSS and responsive telemetry streaming.
- Profiling & Compliance: Evaluated with
adtc-profilerto enforce thermal stability, memory constraints, and reproducible provenance.
🏆 Accomplishments We're Proud Of
- Sub-Millisecond Life Support: Achieving a 0.013 ms fast-path reflex that catches contraindications before an LLM can even complete its prefill stage.
- True Air-Gap Viability: Packaging a full medical synthesis pipeline into an active 2.1 GB RSS footprint, allowing the software to run indefinitely on secondhand laptops hooked up to 15W solar panels.
- Cross-Disciplinary Rigor: Merging deep computational pharmacology with quantized transformer architectures to eliminate medical hallucinations at the edge.
🔮 What's Next for SURYA KARTHI
- Pulse Oximeter Serial Port Bridge: Direct USB/Serial hardware integration with low-cost pulse oximeters for continuous zero-click telemetry ingestion.
- Multi-Dialect Audio Synthesis: Integrating offline, edge-quantized multilingual text-to-speech for local African dialect audio prompts (Swahili, Hausa, Yoruba, Amharic).
- Ultra-Low Power Micro-NPU Deployment: Compiling the Fast-Path reflex engine down to ARM Cortex-M microcontrollers for sub-$5 wearable baby monitoring bands.
