M S Raaghavan

I build NLP systems, LLM applications, natural-language interfaces, graph models, multimodal fusion models, diffusion transformers and distributed data pipelines.

Research

Three peer-reviewed papers on fusing graphs with language.

A crystal lattice passes messages between atoms, text tokens cross-attend, and the band gap is read out.

Nov 20252nd author of 5

MatFusionNet

Transformer–GNN Fusion for Zero-Shot Band Gap Prediction

A multimodal architecture that fuses a six-layer message-passing GNN over periodic crystal graphs with a Transformer text encoder through gated cross-attention.

0.271 eVMAE on the Materials Project benchmark
84,312Compounds evaluated
−7.2%Error vs MEGNet · p < 0.01 over 5 seeds

ICMLDE 2025 · Procedia Computer Science (Elsevier)

Read paper
A URL’s metadata graph lights up, then its graph and text embeddings align.

Jun 20253rd author of 5

GLaMP

Graph-Latent Alignment for Malicious Phishing Detection

A Graph Attention Network over a heterogeneous metadata graph (WHOIS, TLS certificates, redirect chains, threat feeds) with learnable exponential-decay edge weights, aligned to a LoRA fine-tuned LLaMA-7B encoder through an InfoNCE contrastive objective.

91.3%Accuracy
0.90AUC
+4.6 ptsOver a GCN + BERT late-fusion baseline

IEEE ICCCNT 2025 · IIT Indore

Read paper
A residue graph rotates, a partner protein docks, and the annotation is written.

20253rd author of 5

SLPIM

Structural-Linguistic Protein Interaction Modelling

SE(3)-equivariant GNNs over residue proximity graphs, integrated with fine-tuned LLaMA 2 and BioBERT/BioGPT encoders across PDB, STRING, UniProtKB and BioGRID.

90.36%Annotation accuracy
0.56 / 0.48BLEU-4 / ROUGE-L
+13%Over single-modality baselines

CONFLUENCE 2025 · Springer LNCS

Read paper

Projects

Systems I designed, trained and shipped end to end, each with a live demo.

Points stream in, a quadtree partitions them, then a range query and a k-nearest-neighbour search run.

2026Apache Beam · Python · Flask · Leaflet

GeoBeam

Distributed spatial query engine

12.7MNYC taxi trips
4Spatial operations
3Partitioning strategies
  • Range queries, spatial joins, k-nearest-neighbour search and partition counts, built as composable PTransforms on Apache Beam.
  • Pluggable uniform-grid, depth-4 quadtree and reservoir-sampled adaptive-grid partitioning, evaluated over 12.7M NYC TLC taxi trip records.
  • Recovered point geometry from zone IDs via TLC shapefile joins and EPSG:2263 → WGS 84 reprojection; served through a Flask REST API and a Leaflet front end.
Live demoCode
Three faces denoise from pure noise, each steered by different attributes.

2026PyTorch · Diffusion Transformers · GANs · ONNX · WebGPU

HumanDiT

Controllable face generation in the browser

3.78FID on 50k samples
183MParameters
0.54 sPer face on a laptop GPU
  • Trained a 183M-parameter diffusion transformer with flow matching in a frozen DINOv3 latent space (RAEv2) on FFHQ: 256×256 faces steerable by 12 attributes, 89.9% attribute-control accuracy.
  • Distilled it into a single-step GAN student (FID 20.0 vs 23.0 for a flow-distillation baseline) after diagnosing mode collapse and a discriminator noise-schedule bug; audited memorisation against all 70k training faces.
  • Runs in the browser with ONNX Runtime Web and WebGPU, quantised to 4 and 8 bits.
Live demoCode

Experience

Where I learned to turn data into decisions.

Aug 2024 – Jan 2025Kerala

Dev Dungeon

Data Analyst Intern

  • Built interactive Streamlit dashboards that put key business metrics in one self-service view for stakeholders, replacing manual spreadsheet reporting.
  • Analysed customer datasets in Python and Pandas to surface usage and retention trends, giving the team insights that shaped product prioritisation.
  • Contributed to the startup's AI and data science initiatives, supporting its product direction and strategy.

Skills

The tools behind the research and the projects.

LLM & NLP

LLaMABioGPTBioBERTLangChainRAGFAISSChromaDBSentenceTransformersOpenAI API

Modelling

GNNs (GAT, SE(3)-equivariant)TransformersDiffusion modelsGANsComputer visionCross-attention fusionContrastive learning (InfoNCE)LoRAQuantisation

Deep learning

PyTorchPyTorch GeometricTensorFlowHugging Face Transformersscikit-learnNumPyPandasONNX Runtime

Data & distributed systems

Apache BeamBatch & streaming pipelinesSpatial partitioningParquetpyarrowREST API design

Languages

PythonJavaCJavaScriptSQLBash

Tools & infrastructure

GitDockerLinuxGCPFlaskStreamlitMulti-GPU training (A100)Weights & BiasesMySQLMongoDB

Education

Computer science at Amrita University.

Aug 2023 – Present

Amrita University

B.Tech, Computer Science and Engineering

Degree progress—

Relevant coursework

Machine LearningArtificial IntelligenceNatural Language ProcessingGraph MiningData Structures & AlgorithmsDBMSOperating Systems

May 2023Hyderabad

Sri Chaitanya Junior College

Class 12 (MPC)

Jul 2021Hyderabad

Hyderabad Public School

Class 10 (ICSE)

2025 · Leadership & outreach

Lakshya, Student Social Responsibility Program

Jointly delivered a three-hour career guidance and AI ethics workshop to Class 9 and 10 students in Oachira, Kerala.

Read the report