TRACK 1 — ARTIFICIAL INTELLIGENCE

AI

Explore. Automate. Predict. Transform.

Build intelligent solutions with AI, ML, GenAI, NLP, computer vision, and agents — systems that learn from data, automate decisions, and transform how products think.

01SNH-AI-01

CausalTwin: AI That Discovers Why a System Is Failing

Problem Statement

Most AI systems are excellent at predicting what will happen, but significantly weaker at determining why it happened.

In complex environments such as manufacturing, cloud infrastructure, energy systems, logistics, and enterprise operations, dozens of variables can change simultaneously. A conventional ML model may identify that a failure is likely, but it cannot reliably distinguish between:

“Sensor X is correlated with the failure”

and

“Changing Sensor X would actually prevent the failure.”

Challenge

Build a Causal Digital Twin AI system capable of continuously constructing and updating a causal representation of a dynamic system.

The system should:

Ingest heterogeneous time-series and event data.
Discover potential causal relationships.
Separate correlation from intervention-based causality.
Simulate “what-if” interventions.
Predict the downstream effect of an intervention.
Recommend the smallest intervention capable of preventing or mitigating the detected failure.

Example

Instead of:

“Machine temperature indicates 82% failure probability.”

The system should attempt:

“Reducing parameter X by 7% is predicted to reduce failure probability by Y%, while changing parameter Z produces negligible improvement.”

Expected Output

Working causal AI engine
Dynamic causal graph
What-if simulator
Intervention recommendation engine
Confidence/explanation layer
Demonstration using simulated or real dataset

02SNH-AI-02

NeuroRoute: Self-Optimizing AI Inference Architecture

Problem Statement

Modern AI systems frequently use a fixed model architecture:

InputModelOutput

But different inputs require dramatically different computational effort.

A simple query may unnecessarily consume a large model, while a difficult query may require multiple models, tools, retrieval systems, or multimodal reasoning.

Challenge

Build an adaptive AI inference system that dynamically designs its own computational pathway for every input.

Instead of selecting only a model, the system must dynamically determine:

Which model should run?
Whether retrieval is necessary
Whether vision/audio processing is required
Whether an external tool should be invoked
Whether another model should verify the result
How much computational budget should be allocated
When the system should stop reasoning

Required Architecture

InputComplexity EstimatorModel/Tool Graph GenerationExecutionConfidence CheckEscalation/Termination

Example

Easy requestSmall model120 ms

Moderate requestRAG + medium model600 ms

High-risk requestMultiple models + verification2.5 sec

Expected Output

Teams should demonstrate:

At least 3 computational components/models
Dynamic routing
Confidence estimation
Cost/latency tracking
Automatic escalation
Measurable improvement over a fixed pipeline

03SNH-AI-03

MemoryMesh: Conflict-Aware Long-Term AI Memory

Problem Statement

AI agents increasingly maintain memories about users, organizations, tasks, and environments.

But conventional memory systems generally treat memory as:

StoreRetrieveGenerate

Real-world memory is much harder.

Information becomes:

Outdated
Contradictory
Duplicated
Context-dependent
Partially reliable
Temporally invalid

Challenge

Build an AI memory system capable of determining whether a new piece of information should replace, modify, coexist with, or reject an existing memory.

The system should maintain:

Temporal validity
Source provenance
Confidence
Contradictions
Context
Dependency relationships

Example

Memory 1:

“Server A uses Python 3.10.”

Later:

“Server A migrated to Python 3.12.”

The system shouldn't simply append another memory.

It should understand:

Python 3.10historically valid

Python 3.12currently valid

Expected Output

Memory graph
Contradiction engine
Temporal reasoning system
Provenance tracker
Context-aware retrieval
Demonstration involving intentionally conflicting information

04SNH-AI-04

SenseFusion: AI for Missing-Sensor Intelligence

Problem Statement

Real-world AI systems rarely have perfect sensor availability.

Cameras fail.

Microphones become noisy.

IoT sensors disconnect.

Network packets disappear.

Instead of simply reporting:

“Sensor unavailable.”

build an AI system capable of reasoning about the missing information itself.

Challenge

Develop a multimodal AI system capable of dynamically reconstructing missing information from available modalities.

For example:

Camera unavailable + microphone + vibration + historical context

→ infer environmental state.

But the system must distinguish between:

Observed information

and

AI-inferred information.

Required Capabilities

The system should:

Detect missing or degraded modalities.
Estimate reliability of remaining modalities.
Generate a temporary representation of missing information.
Quantify uncertainty.
Continue operation without falsely presenting inference as observation.
Recalibrate when the missing sensor becomes available again.

Expected Output

A working prototype demonstrating at least three modalities, intentional sensor failure, adaptive reconstruction, and uncertainty visualization.

05SNH-AI-05

TrustChain AI: Self-Auditing AI Decision Engine

Problem Statement

AI systems can generate answers and decisions, but determining how trustworthy a specific decision is remains difficult.

Build an AI system that does not merely produce:

Decision: APPROVE

but produces an internal technical audit trail:

Evidence usedreasoning pathuncertaintyconflicting evidencepolicy constraintsdecisionpost-decision outcome.

Challenge

Develop an AI decision engine capable of automatically auditing its own decisions and learning from post-decision outcomes.

The system should:

Identify evidence used
Track source provenance
Detect conflicting evidence
Estimate confidence
Identify missing information
Generate a machine-readable decision trace
Compare predicted outcome with actual outcome
Update future decision confidence

Expected Output

A working AI decision system with:

Decision engine
Evidence graph
Confidence mechanism
Audit trail
Outcome feedback
Error attribution