Home / Research Themes

Research Themes

The lab operates six specialized student-led research pods. Each pod is anchored by senior student leads, governed by cross-pod councils, and committed to open computational reproducibility.

01
View Pod Spec

Conducting applied research on low-resource Indic language tokenization, Kannada sentiment and dialect parsing, cross-lingual pseudo-labeling, and compute-efficient parameter-efficient fine-tuning (PEFT/LoRA) on open weights. Our objective is establishing verified linguistic benchmarks that serve under-resourced Dravidian dialects with rigorous evaluation criteria.

Core Research Questions

  • How do byte-pair encoding vocabularies compare with character-level and morphological tokenizers on agglutinative Dravidian syntax?
  • Can cross-lingual teacher-student distillation retain high confidence bounds under severe domain shift?
  • What are the minimal parameter ranks for LoRA adapters to achieve parity with full fine-tuning on regional sentiment classification?
Methods
Morphological parsing and boundary analysisTeacher-student cross-lingual knowledge distillationParameter-Efficient Fine-Tuning (LoRA/QLoRA)
Tech Stack
PyTorchHugging Face TransformersIndicBERTspaCyWeights & Biases
02
View Pod Spec

Designing latency-critical perception pipelines, INT8/FP4 quantization schemes, multi-camera telemetry streaming, and embedded hardware acceleration across NVIDIA Jetson micro-architectures. We study empirical tradeoffs between frame latency, thermal throttling, and precision degradation on resource-constrained platforms.

Core Research Questions

  • What is the empirical Pareto frontier between INT8 calibration loss and inference FPS on embedded edge GPUs?
  • How can zero-copy CUDA memory layouts prevent frame bottlenecks in multi-stream RTSP ingestion?
  • What architectural modifications allow YOLO-family backbones to maintain mAP under aggressive low-bit integer quantization?
Methods
Post-Training Quantization (PTQ) & Quantization-Aware Training (QAT)Hardware-in-the-loop thermal and power telemetryTensorRT engine layer fusion and kernel autotuning
Tech Stack
TensorRTYOLOv8CUDAOpenCVONNXDeepStream
03
View Pod Spec

Formulating formal state-machine specifications, deterministic consensus protocols, and verifiable guardrails for autonomous agent systems. We focus on eliminating non-deterministic failure modes and unconstrained recursive tool loops in multi-agent software engineering pipelines.

Core Research Questions

  • Can deterministic state transitions guarantee termination in arbitrary multi-agent tool execution graphs?
  • What attribution scoring methods best measure hallucination rate reductions in source-grounded RAG architectures?
  • How do distributed consensus voting protocols across heterogeneous local LLMs compare to centralized monolithic prompting?
Methods
State-machine graph modeling with LangGraphSentence-level citation and graph attribution validationSchema-enforced JSON schema and Pydantic validation boundaries
Tech Stack
LangGraphFastAPIOllamaPydanticDockerChromaDB
04
View Pod Spec

Investigating lightweight neural network architectures for real-time biological signal classification and clinical telemetry. Our work focuses on patient data privacy, artifact suppression in noisy portable sensor records, and deterministic edge classification of cardiac anomalies.

Core Research Questions

  • Can ultra-compact 1D-CNN models match residual attention networks on single-lead arrhythmia classification?
  • What wavelet thresholding parameters best preserve spectral EEG power while eliminating ocular blink artifacts?
  • How do models trained on clean clinical benchmarks degrade when presented with consumer-grade wearable noise?
Methods
Continuous Wavelet Transform (CWT) and discrete wavelet decomposition1D Temporal Convolutional Networks with depthwise separable filtersInter-patient cross-validation on PhysioNet MIT-BIH datasets
Tech Stack
PyTorchWFDBNeuroKit2SciPyMNE-Python
05
View Pod Spec

Auditing adversarial threat vectors in open-weight language models and measuring demographic parity in automated evaluation pipelines. We build standardized red-teaming test suites mapped to OWASP and MITRE ATLAS security frameworks to detect jailbreaks, indirect prompt injection, and bias transfer.

Core Research Questions

  • Which prompt injection attack vectors reliably bypass system prompt boundaries across quantized 7B-parameter models?
  • How can disparate impact metrics be baked into continuous integration pipelines for tabular screening heuristics?
  • What verifiable watermarking schemes survive temperature sampling and paraphrase attacks?
Methods
Automated adversarial prompt fuzzing and boundary probingDemographic parity and equalized odds statistical auditingModel card and dataset lineage validation protocols
Tech Stack
GarakMITRE ATLASAIF360FairlearnSHAPPyTorch
06
View Pod Spec

Engineering research infrastructure, automated reproducibility testbeds, and distributed GPU profiling tools for lab members. We enforce reproducible experiment tracking where model weights, seed randomness, hardware memory bounds, and code revisions are immutably logged for every pull request.

Core Research Questions

  • What minimal container telemetry capture is required to guarantee bit-level replication of CUDA kernels?
  • How can multi-tenant GPU memory fragmentation be minimized during automated test execution?
  • What network bandwidth thresholds govern model partition splits across heterogeneous edge-server environments?
Methods
Prometheus GPU scrape exporters and VRAM peak trackingAutomated container build runners with pinned dependency hashesHeterogeneous RPC model layer offload benchmarking
Tech Stack
DockerMLflowGitHub ActionsPrometheusTriton ServerKubernetes