Manini Banerjee

I’m a Design Engineer building machines that sense, learn, and adapt, often by borrowing from biology.

Currently: Master of Design Engineering @ Harvard

All Robotics & Hardware ML & Sensing Climate & Bio-design

COMPUTATIONAL ECOLOGY:

BIOPOD Co.
ECOLOGY · INFRASTRUCTURE · SYSTEMS
Floating mycelium modules that restore wetlands and clean waterways, deployed in USA & India

Ecological AI

PREDICTION · INTERFACE · DATA
A decision-support tool that predicts where ecosystem restoration will have the most impact.

Algorithmic Morphogenesis

BIO-COMPUTATION · DATA MATERIALIZATION
Live EEG brain signals shape the growth of algae, using light to steer how it moves.

HARDWARE & INTERFACES:

Threads
EDGE ML · HARDWARE · TEXTILES
Textiles that sense touch and run machine learning on-device.

S(kin)-orb
HAPTICS · BIOSENSING · AFFECTIVE COMP. A device that turns muscle signals (EMG) into touch, so two people can feel each other’s emotions remotely.

Vermiform

COMPONENT · SOFT ROBOTICS · WEARABLE
Soft robotic components for wearables, modeled on biological forms.

Chito-bot
BIOCOMPOSITES · TRANSIENT ELECTRONICS
A six-legged robot built from biodegradable composites, testing how bio-materials hold up structurally.

Siphonophorae textilis
SOFT ROBOTICS · BIOHYBRID · TEXTILES
A slow-swimming biohybrid robot whose knit fins host metal-capturing microbes, cleaning heavy metals from mining wastewater.

STRATEGIC SYSTEMS:

PFV

MOBILITY · ECOLOGY · SYSTEMS

Autonomous vehicles redesigned as urban ecological infrastructure with Hyundai Motor Group R&D.

Aero

SENSING · MATERIALS · DATA
Robotic materials that sense local air quality and capture pollutants.

Bio - intelligence
COMPUTATION · SYSTEMS · BIOLOGY
Exploring living organisms as an alternative to digital control systems.

Ecological AI

Interface & Prediction

A Decision-Support System for Targeted Ecosystem Restoration.

PROBLEM
Restoration is often paralyzed by complexity.

Individuals struggle to cross-reference thousands of plant species against site-specific constraints (salinity, pH, wave energy) to find viable solutions.

RESPONSE

Ecological AI is a web-based digital twin that translates complex biodiversity data into actionable physical infrastructure. It functions as a parametric configurator, allowing non-experts to generate scientifically validated "restoration kits" tailored to their local environment.

TECH STACK:

FRONTEND

Frameworks: Flask (Python Web Server), JavaScript (ES6), HTML5/CSS3.

Morphology: A custom-built geometry configurator that allows users to manipulate the physical form of the pod (Stretch, Inflation, Lattice Density) using Paul Bourke’s Supershape formula

Logic: Users can optimize for buoyancy and root-attachment surface area. The interface provides real-time visual feedback on how the physical constraints interact with the biological goals.

BACKEND

Pipeline: Python (Pandas, Numpy) processing of the USDA Plants Database and National Wetland Plant List (NWPL).

Decision Engine: "Evidence-Weighted" filtering algorithm that cross-references site conditions (Salinity, pH) against plant tolerance data to generate valid "Restoration Kits."

Integration: The backend feeds valid plant species into the frontend 3D model, ensuring the digital twin represents a biologically viable ecosystem.

TECHNICAL PIPELINE

"Evidence-Weighted" Ensemble

Layer 1 - Hard Constraints - Filters for strictly impossible survival scenarios using NWPL wetland indicators and USDA tolerance data.

Layer 2 - Goal Ranking - Applies weighted scoring to prioritize species based on specific restoration targets (e.g., nitrogen fixation vs. bank stabilization).

Layer 3 - Evidence Modifiers - Adjusts confidence scores based on real-time field inputs (salinity, pH) and future iNaturalist validation layers.

METHODOLOGY

Bias-Aware Modeling: Explicitly flags "unknowns" instead of interpolating false certainty.

Data Provenance: All datasets accessed programmatically via reproducible notebooks.

The 3D form generation utilizes Paul Bourke’s Supershape Formula to drive parametric geometry.

Specific parameters (m, n1, n2, n3) controls surface-area-to-volume ratios. This ensures that every 'aesthetic' change in the slider directly correlates to a performance metric (e.g., increased root filtration surface).

Form Follows Performance: Users manipulate sliders ("Stretch," "Inflate") to optimize the pod’s physical properties.

Biological Logic: Increasing the "Stretch" parameter maximizes surface area for root intermingling and filtration efficiency, directly linking code to ecological function.

FIELD VALIDATION & ITERATION!!

Context: Deployed for A/B testing at the Cambridge Science Carnival (100+ unique interactions).
Insight: Users struggled with the sensitivity of the "Inflation" slider, often creating non-viable geometries.
Action: V2 Roadmap includes "Elastic Snapping" on sliders to guide users toward biologically valid constraints while maintaining agency.