Why Fine-tuning Beats Prompting for Satellite AI
Everyone reaches for prompt engineering first. For general language tasks, that's often enough. For satellite imagery, it almost never is — and here's why the gap exists.
Apr 10, 2026
Open to new opportunities
Tushar Thokdar
01.
How I Think
Most satellite AI projects stall between research and reality — models that work in notebooks but never reach production. I focus on closing that gap: taking a 300M parameter foundation model, adapting it to a real earth observation task, and delivering 4x the baseline performance with docs, inference code, and a pipeline anyone can reproduce.
I work across the full ML stack — satellite vision, Gen AI, RAG pipelines, agentic workflows — because most hard problems don't fit neatly into one discipline. The goal is always a working system, not just a trained model.
At Godel Space I built production satellite computer vision pipelines for disaster monitoring. At Unify I solved LLM provider fragmentation with a modular integration framework. Won 1st place at an edge AI hackathon by deploying an optimized model at 4x the speed with 85% of the accuracy.
LocationRemote-first, Kolkata
TimezoneIST (flexible overlap)
Work styleShip fast, document well
AvailabilityImmediate
02.
Problems I Solve
Each of these maps to a real challenge I've worked through — from raw data to a running system.
Satellite AI Models
Pre-trained models don't understand satellite data out of the box. I fine-tune foundation models (Prithvi, AnySat) on domain-specific imagery to make them actually useful for detection, segmentation, and monitoring.
Prithvi EO-2.0AnySatUPerNetViTJEPASentinel-1/2
End-to-End ML Pipelines
Most ML work stops at the notebook. I build the full loop — data collection, preprocessing, training, evaluation, and deployment — so the model actually runs in production.
PyTorchPyTorch LightningHuggingFaceTensorFlowscikit-learn
Geospatial Data Engineering
Satellite data is noisy, multi-spectral, and massive. I handle acquisition and processing at scale — multi-sensor fusion, temporal stacking, spectral analysis — so models have clean inputs.
Google Earth EngineGDALRasterioGeoPandasOpenCV
Production Deployment
A model no one else can run solves nothing. I ship with proper docs, inference APIs, containerized environments, and model versioning so teams can actually use and maintain what I build.
DockerFastAPIHuggingFace HubGCPONNX Runtime
Gen AI & LLM Applications
Off-the-shelf LLMs hallucinate on domain data. I solve that through fine-tuning, RAG pipelines, and agentic workflows — building products that are reliable in the real use case, not just demos.
LLM Fine-tuningRAG PipelinesAgentic AILangChainChatbotsPrompt Engineering
Backend & Data
AI systems need reliable APIs and data infrastructure around them. I build the backend layer — APIs, data processing, databases — that holds the whole thing together.
PythonFastAPIFlaskPostgreSQLConvexNumPyPandas
03.
Problems Solved
Each project started with a specific gap or failure mode. Here's what the problem was and what came out of it.
Shipped
Prithvi-2.0 Flood Detection
Flood mapping from satellite imagery is critical for disaster response, but pre-trained models perform poorly out of the box. Fine-tuned Prithvi EO-2.0 on Sen1Floods11 to produce accurate, production-ready flood segmentation.
0.72 IoU+437% vs Baseline
Flood Detection IoUPrithvi-2.0UPerNetPyTorchSentinel-2+1
Research
Wildfire Burn Scar Detection Pipeline
Wildfire burn scar mapping requires understanding temporal change, not just a snapshot. Built an end-to-end pipeline with a novel Delta Channel Algorithm to explicitly encode pre/post-fire spectral differences, significantly improving classification accuracy.
+50.58%Novel Algorithm
Macro F1 ImprovementGEEPrithvi-2.0PyTorch LightningSentinel-2+1
Research
AnySat Multi-Sensor Flood Inference
Optical imagery alone fails in cloudy conditions — a real constraint for flood detection. Implemented AnySat (CVPR 2025) to fuse SAR and optical data from Sentinel-1 and Sentinel-2, enabling all-weather flood segmentation.
125MCVPR 2025
Parameters (Multi-Modal)AnySatJEPASAR FusionSentinel-1/2+2
04.
Where I've Worked
AI Engineer (Contract)
Godel Space
Dec 2024 — Feb 2026
Remote
Problem: satellite imagery for disaster monitoring requires multi-sensor fusion at scale — a hard data engineering and ML challenge. Built the full pipeline from raw Sentinel data to deployed inference.
- ▸Shipped AnySat multi-modal inference pipeline processing Sentinel-1/2 satellite data
- ▸Built preprocessing pipelines for SAR-optical imagery fusion at scale
- ▸Created reproducible PyTorch training scripts with experiment tracking
- ▸Explored FastAPI for model serving APIs
Result
Delivered production-ready multi-sensor satellite inference system for disaster monitoring.
AI Engineer (Contract)
Unify
Aug 2024 — Dec 2024
Remote
Problem: teams building with LLMs were locked into single providers, making cost/quality trade-offs impossible. Built a modular integration layer so switching providers or running A/B tests required minimal code changes.
- ▸Integrated GPT-4, Llama, and Claude APIs into prototype applications
- ▸Built content filtering and guardrails reducing harmful outputs
- ▸Created modular API architecture for comparing LLM providers on quality and cost
- ▸Shipped prompt engineering tools and RAG pipelines
Result
Delivered production-ready LLM framework enabling instant provider switching and A/B testing.
05.
Recognition & Background
🏆
1st Place — AI Startup Hackathon
Edge Runners
The challenge: deploy a capable AI model on hardware with strict memory and compute limits. The solution: fine-tuned Phi-3 with quantization and pruning, hitting 85% of full-model accuracy at 4x the speed — good enough to be genuinely useful at the edge.
Edge AIModel OptimizationDocker
Education
8.6
B.Tech Information Technology
Haldia Institute of Technology · 2020-2024
Certifications
Python Programming Essentials
Rice University
A Crash Course in Data Science
Johns Hopkins University
Complete SQL Bootcamp
Udemy
06.
All postsWriting
Notes on satellite AI, edge deployment, and things that broke before they worked.
07.
Let's Work Together
Ready to start immediately
If you have a hard problem in satellite AI, earth observation, or ML systems — or a role where that kind of thinking is useful — I'd love to hear about it.