Reacher AI

Inspiration

Healthcare access in India is uneven. 67% of Indians live 5+ km from a health facility, only 1 in 5 people with mental illness ever receives treatment, and rural India has about 3 psychiatrists per 1,00,000 people. Emergencies like chest pain, stroke, and suicidal ideation are often missed because people rely on informal self-triage, and a delayed first assessment for something like skin cancer can change the outcome.

We asked a simple question: what if the first screening didn't require a clinic, a smartphone data plan, or any money? Most people in underserved areas already have a basic phone. Reacher AI meets them there, over SMS, voice, and a smartphone camera, at ₹0 cost to the patient.

What it does

Reacher AI is a unified AI health screening platform with three modules behind a safety-first gate:

TriageBot: instant emergency triage and care routing over SMS. It covers 20 symptom clusters in English and Hindi, keeps stateful sessions, and works on a 2G connection.
MindCheck: conversational and voice mental health screening based on the validated PHQ-2 and GAD-2 instruments, which WHO's mhGAP guide endorses for community health worker use.
ScanCheck: smartphone-camera skin screening with an image quality gate (sharpness, brightness, and contrast checks) so poor photos are rejected before analysis.

Before any AI module runs, a red-flag engine checks for emergencies. If triggered, the patient is directed straight to 112 / Tele-MANAS, with no LLM involved. Safe cases are assigned a risk band (Emergency, Urgent, Routine, or Self-Care) and surfaced to a clinician dashboard where they can be reviewed, acknowledged, and acted on.

Every output states that it is an AI screening aid, not a diagnosis.

How we built it

Safety engine: pure Python rule engine, under 1 ms latency, tested with 16/16 passing test cases.
Backend: Python APIs with SQLAlchemy persistence (SQLite now, Postgres-ready) and Pydantic to enforce valid input ranges on the clinical screeners.
SMS: a Twilio webhook endpoint for TriageBot, built and tested.
Skin screening: a real image quality gate feeding a CNN inference pipeline. The pipeline is real, but the model is in demo mode until it is trained on ISIC and APTOS data.
Dashboard: a React single-page app talking to the live API, with a mock fallback and an audit log.
Deployment: CPU-only, no GPU required, and built so the whole system can be demoed offline from one laptop.

Challenges we ran into

Making emergencies fast and trustworthy. We kept the red-flag path free of any LLM so it is deterministic, instant, and testable.
Designing for low connectivity. The SMS triage had to work on 2G, and every module needed a fallback so the system still functions without a persistent internet connection.
Dataset bias. The ISIC archive is skewed toward lighter skin tones (Fitzpatrick I-III), and APTOS has a domain gap between fundus cameras and smartphones. Both limit how well a model would perform for South Asian patients.
Staying within scope. We kept the product to screening and routing with no diagnosis or treatment claims, which also keeps it within the CDSCO wellness/health-information category for an MVP.

Accomplishments that we're proud of

A working end-to-end system: safety engine, triage classifier, SMS webhook, mental health screener, image quality gate, database, and clinician dashboard.
A safety-first architecture where emergencies never depend on an LLM.
Clinically grounded screening using validated instruments (PHQ-2 and GAD-2) with published sensitivity and specificity.
A realistic cost model: about ₹4,000-₹6,000/month to host and about ₹1,500/month for a 10,000-message pilot.
Honesty about limits: we clearly mark the CNN as demo mode and document dataset bias instead of overclaiming.

What we learned

In health tools, safety-critical paths should be simple, deterministic, and testable, and AI belongs behind them rather than in front.
Accessibility is a design constraint, not an afterthought. SMS and offline-first choices shaped the whole architecture.
Validated clinical instruments give a stronger foundation than inventing our own screening questions.
Data representation matters as much as model quality, especially for skin screening across diverse skin tones.

What's next for Reacher AI

Train the ScanCheck model on ISIC and APTOS, and add smartphone-native data (such as PAD-UFES-20) and South Asian skin-tone data to close the bias gap.
Run a pilot using the 10,000-message SMS setup to measure real-world reach and referral outcomes.
Partner with community health workers (ASHAs and ANMs) so Reacher AI works as a triage aid in the field.
Move from SQLite to Postgres for production scale.
Scale toward the target population of about 300 million rural Indians with basic mobile connectivity.