ML systems · scientific computing · backend engineering

Shaunak Rane   /   Applied ML & systems

Intelligence,
put to work.

Machine-learning systems engineer

I build validated ML systems that turn complex physical and operational data into reliable decisions across industrial diagnostics, agent orchestration, graph learning, and distributed energy.

Start a conversation
Selected work

Four systems. Different scales. One engineering mindset.

01 / Agent systems

Multi-agent swarmGemini 2.5 GatewaySupabase pgvector

Aegisthreat intelligence

An investigation platform that turns noisy claims into traceable, evidence-backed verdicts.

Agent Swarm2-stage decoupled
Market MLZ-score anomalies
RSS Ingestion15 min loop
Case study Live System

Production engineering: serverless edge key cycling, structured JSON Schema contracts, and pgvector semantic recall.

Autonomous investigation system combining SHA-256 claim deduplication, quantitative market anomaly detection (Z-scores), and 15-minute RSS ingestion with an asynchronous two-stage Gemini verification swarm.

02 / Energy systems

Continuous RLDePIN Architecturezk-SNARKs

Gridiumprotocol

A simulated energy market where learning agents coordinate local power and respond to network imbalance.

500msphysics tick rate
7D → 1DDDPG continuous fee
15Ohm's law nodes
Case study Live Website

Closed-loop stability: DDPG steers AMM swap fees dynamically to avert localized microgrid brownouts.

A distributed microgrid economy pairing a 15-node Ohm's law physics simulator with a continuous-control PyTorch DDPG agent, 500ms Socket.io gateway, constant-product EVM AMM, and Groth16 zk-SNARK settlement.

03 / Industrial intelligence

AI internshipIndustrial MLFlask

Compressorevidence system

Turning compressor operating data into evidence a maintenance team can review.

3plant-specific contracts
3,879DFI rows reproduced
17/17validation gates
Full field report

Forecasting remains an experimental visualization and never drives maintenance alerts.

A condition-monitoring platform that separates what the thermodynamics support from what a polished forecast merely suggests.

04 / Graph intelligence

Independent studyGraph MLPhysics benchmark

TopoFlowpore network

Learning how fluid moves through rock, and testing when graph models improve on a physics baseline.

1,231pore-network graphs
5rock families tested
46.2%best MSE reduction
Technical case study

Benchmark result: learned topology helps in heterogeneous media; it is not a universal replacement for physics.

A permeability study where the topology is learned, the classical baseline stays visible, and pore heterogeneity decides which model deserves trust.

05 / Beyond the projects

What I bring
to a team

I move comfortably between investigation and implementation: data forensics, ML evaluation, APIs, Docker, responsive interfaces, and the documentation that keeps claims honest.

AI InternFlask REST services, thermodynamic feature engineering, and pytest validation for a compressor CBM platform.
Software InternFastAPI backend, BOM pipeline, and Random Forest attendance prediction for food operations.
Independent workTopoFlow GNN, Aegis, and Gridium across backend, ML, and distributed-system design.

Have a meaningful problem to solve?

Let’s build
something useful.

FIELD REPORT

Evidence

PROJECT INDEX / 06 SYSTEMS

Archive