HackDays Delhi
Theme: Beneath the Water

Every dive,
watched over.

Abyss.ai is a vision + telemetry command dashboard for autonomous underwater vehicles. It fuses Gemini's multimodal reasoning with live pressure telemetry to detect, score and explain subsea infrastructure risk the moment it appears — not hours later, back on the surface.

Gemini 3.6 Flash is the reasoning engine, not a bolt-on.

Risk fusion — live instrument
R(t)

R(t) = (α · Cgemini) + (β · |Pnow − Pbase| / Pbase)

Gemini confidence

0.82

Corrosion, frame #4471

Pressure deviation

4.6%

vs. baseline at depth

Computed risk score0.65 · Elevated

α = 0.7, β = 0.3 — operator-tunable defaults, not values calibrated against real damage data.

For judges

30-second orientation

Working demo

Sign up, sign in, and the command dashboard is live — not slides.

Full source code

Public repository

Gemini 3.6 Flash core

The reasoning engine driving every anomaly score, not a bolt-on feature.

Fastest way to evaluate: sign in, click “Live Judge Test” in the dashboard, upload any photo of underwater damage or corrosion from your phone, and watch a real Gemini API round-trip return structured JSON in front of you.

The problem

Subsea monitoring is slow, expensive, and human-bottlenecked

Monitoring deep-sea infrastructure today means AUVs surfacing, hard drives being physically extracted, and hours of manual video review before anyone knows if there's a problem. Abyss.ai's automated frame-by-frame analysis flags critical anomalies the moment they appear.

Estimates based on published industry review ratios, not a customer-validated figure.

Manual review cost

~$30,000/vessel-day

Review ratio

4h : 1h

human review per hour of footage

Target reduction

~95%

in review time

Meaningful AI integration

Gemini 3.6 Flash reads the ocean, not just text

Image buffers — camera frames and sonar spectrograms — stream directly into Gemini's multimodal vision pipeline, prompted to act as a Subsea Infrastructure Analyst.

  1. 01

    Frame capture

    AUV camera frames and sonar spectrograms stream into the pipeline as raw image buffers.

  2. 02

    Gemini vision

    Prompted as a Subsea Infrastructure Analyst, Gemini 3.6 Flash reads each frame for corrosion, leaks and ecological damage.

  3. 03

    Strict JSON schema

    The model returns a typed anomaly category, confidence score and explanation — no unstructured text to clean up.

  4. 04

    Risk fusion

    Visual confidence is fused with live pressure telemetry into a single operator-facing risk score.

Sensor fusion

Don't trust vision alone

Gemini's visual confidence is fused with live physical telemetry — pressure deviation from the expected baseline at depth — so a visual anomaly is contextualized against what the hardware is actually experiencing.

α and β are exposed as live draggable sliders in the dashboard's Settings panel, so an operator can weight visual vs. physical signal differently depending on mission type.

Variables

Fed into the live risk gauge on every mission

R(t)Computed risk score (0.0–1.0)
CₑₑₘᵢₙᵢGemini's visual-damage confidence output
PᴄᴜᴿᴿᴇɴᴛLive ambient pressure telemetry
PᵇᵃˢᵉˡⁱⁿᵉExpected pressure at current depth
α, βOperator-tunable weighting coefficients (0.7 / 0.3 default)

Full transparency

What's real vs. simulated

We'd rather you trust the majority that's real than be impressed by a demo that isn't. Built in 10 hours, here's exactly where the line sits.

Gemini 3.6 Flash vision analysis
Real

Live API call, triggered on demand via “Live Judge Test”

Structured JSON parsing
Real

Enforced schema, feeds dashboard directly

Risk fusion calculation R(t)
Real

Runs live against telemetry + Gemini output

Dashboard UI / auth (Clerk)
Real

Functional sign-in, not a static mock

AUV hardware telemetry
Simulated

No physical AUV in Delhi — local Node.js generator produces realistic pressure/depth streams

AUV camera feed
Simulated

Looped underwater footage stands in for a live feed

Usage-based billing dashboard
Simulated

Local-state UI demonstrating the intended SaaS flow — no live metering or payment wired up

Business model

Built as multi-tenant SaaS, not a single demo app

Clerk handles organization-based sign-in and routing; Supabase Row-Level Security scopes mission logs per tenant, so multiple offshore operators — drilling, telecom — can run isolated missions on the same platform.

Intended pricing: pay-per-mission-hour of footage processed, so operators avoid investing in dedicated on-vessel AI hardware.

If we had more time

Next on the roadmap

  • Wire real Stripe metering to the billing dashboard
  • Replace the looped video feed with a real-time video stream input
  • Validate α/β against a labeled dataset of real subsea damage imagery
  • Load-test the Gemini pipeline against AI Studio's free-tier rate limits

Technical architecture

Built entirely on free-tier infrastructure

LayerTechnologyPurpose
FrontendNext.js 14, Tailwind CSS, shadcn/uiHigh-contrast operator command dashboard
HostingVercel (serverless functions)API routes and deployment
Auth & Multi-tenancyClerkOrg-based sign-in and routing
AI EngineGoogle Gemini 3.6 Flash APIMultimodal visual anomaly analysis → structured JSON
DatabaseSupabase (PostgreSQL)Mission logs, telemetry history, RLS-scoped tenant data
Data StreamCustom Node.js mock generatorSimulates high-frequency AUV pressure/depth telemetry

See what's happening beneath the surface.

Sign up, open the command dashboard, and run a real Gemini analysis on your own photo in under a minute.