Evaluation
Reasoning stress tests
Multi-turn judge harness for plant science, soil, and climate plans with grounded open-ended scoring.
PhD @ Iowa State · Applied AI
Shreyan Ganguly
I evaluate and distill large models so they behave on real, noisy data—whether the inputs are crop images, EEG traces, or speech captured in the field.
- Now Research Assistant · Mechanical Engineering (Iowa State)
- Focus LLM/VLM evaluation, speech + EEG distillation, ag/biomed agents
- Stack Python · PyTorch · Hugging Face · Retrieval
Evaluation first
LLMs/VLMs audited with domain judges for agriculture, and scientific reasoning.
Distillation
Deep Model compression to bring ASR + EEG pipelines to edge hardware.
Field notes
Messy sensors → retrieval + tool use stacks for agronomists & biomedical teams.
Pulse
Three current threads I’m iterating on.
Distillation
Model compression and calibration
Adaptive teachers, latency-aware losses, and calibration audits.
Systems
Retrieval + tools
LLMs that cite agronomy notes, vision cues, and sensor metadata before replying.
Recent work
Short, high-signal blurbs. Ping me for deeper dives.
LLMs · 2025
Large reasoning models for agriculture
Benchmarked top LLM/VLM stacks on multi-step agronomy decisions; exposed gaps in retrieval grounding and temporal planning.
Speech · 2024
AdaKD for ASR compression
Teacher–student distillation with adaptive loss schedules; maintained WER while cutting latency for edge clients.
Signals · 2024
EEG seizure detection
Hybrid CNN/transformer encoders with calibration diagnostics across animals and labs.
Vision
Shot boundary & plant ID
Video forensics tooling for archival footage plus fine-grained plant classifiers tuned for field noise.
Experience
A quick timeline of roles that shaped my research style.
2024 — Present
Research Assistant · Iowa State
Lead eval efforts for large reasoning models; build retrieval + tool use interfaces for agronomy partners.
2023 — 2024
Research Associate · IISc
Distilled streaming ASR models, tuned multilingual evaluation harnesses, and shipped prototypes to edge devices.
Earlier
Internships & open-source
Media forensics, plant phenotyping, and GSoC experiences working on pragmatic CV pipelines.
Toolkit & notes
Tools I lean on and notes I’m writing next.
Stack
Python
PyTorch
Hugging Face
Lightning
Weights & Biases
Vector stores
LLM/VLM evaluation
Knowledge distillation
Prompt + tool orchestration
Notes draft queue
-
Evaluating LLMs for agriculture
Grounded metrics, domain shift, and how to score tool use.
-
Edge-ready ASR
How I decide what to keep vs. compress in speech students.
-
Deepfake reality checks
Designing detectors that survive messy newsroom footage.
Contact
Reach out for internships, collabs, or evaluation chats.
Email works best. Include a sentence on the problem you’re tackling.
Email: shreyang@iastate.edu