About AdaptShot¶
Built in Tanzania for the World¶
AdaptShot is proof that world-class AI can be built anywhere, with any resources, when guided by a clear mission: make AI work for the people who need it most.
🌍 The Origin Story¶
The Problem¶
In 2025, I watched an agricultural extension worker in Mbeya, Tanzania try to diagnose crop disease from phone photos. She had years of training and a smartphone—but every AI tool available to her required cloud connectivity, GPU acceleration, or thousands of labeled images. None of those exist in rural Tanzania.
Every "state-of-the-art" solution failed her.
That moment crystallized a truth: modern AI has been optimized for Silicon Valley, not for the world. 80% of the global population lives in regions with unreliable internet, limited electricity, or both—yet 95% of AI research assumes the opposite.
The Response¶
AdaptShot was built to flip this paradigm. It is:
- CPU-first: Runs on laptops and single-board computers, no GPU required
- Offline-capable: Fully functional without internet after initial installation
- Few-shot: Learns from as few as 5 examples per class
- Memory-efficient: Operates within 250MB RAM indefinitely
- Human-aligned: Every prediction is calibrated and every correction improves the model
🎯 Mission & Vision¶
Our Mission¶
Democratize trustworthy, CPU-first, human-in-the-loop few-shot vision AI for resource-constrained environments worldwide.
Who We Build For¶
| User | Use Case | Why AdaptShot |
|---|---|---|
| 🧑🌾 Small-scale farmers | Crop disease identification from 10 leaf photos | Works offline on low-cost Android phones |
| 🏥 Rural healthcare workers | Triage pneumonia from chest X-rays without a radiologist | Calibrated confidence prevents dangerous overconfidence |
| 🦁 Wildlife conservationists | Classify camera trap images in remote areas | Runs on battery-powered Raspberry Pi |
| 🏭 Quality control inspectors | Detect manufacturing defects from few reference samples | Learns new defect types from human corrections |
| 🎓 Students & educators | Learn AI concepts with real, runnable code | No GPU, no cloud, no expensive hardware |
Core Values¶
- Truth Over Hype — We document what works, what doesn't, and why. No exaggerated claims.
- Constraint-First Engineering — We design for the hardest constraints first. If it runs on a $50 phone in rural Tanzania, it runs anywhere.
- Human Dignity — AI should augment human expertise, not replace it. Uncertain predictions are flagged for review, never silently wrong.
- Environmental Responsibility — CPU-first inference uses 10–100× less energy than GPU alternatives. Carbon-aware configuration lets users optimize further.
- Global South First — We build for the user in Mbeya before the user in Menlo Park. Innovation should serve the many, not the few.
📜 Technical Journey¶
v0.1.0 — The Foundation (May 2024)¶
Frozen backbone extraction (ResNet-18, MobileNetV3-Small) with cosine similarity search. 68% accuracy on CIFAR-10 with 10 images per class at 90ms latency on a consumer CPU. Proved that useful few-shot inference doesn't require GPUs.
v0.1.1 — Trust & Continual Learning (June 2025)¶
Introduced online temperature scaling with sliding-window ECE tracking, Adaptive Confidence Thresholding (ACT), CA-EWC head fine-tuning, and UP-UGF buffer pruning. Added energy profiling, embedding caching, and OOD detection.
v0.2.0 — Production Hardening (Current)¶
Comprehensive algorithmic and production hardening across all subsystems:
- Contrastive Learning: Gradient-trained InfoNCE projection head with 2-layer MLP and full backpropagation
- Conformal Prediction: True leave-one-out prototype recomputation for valid finite-sample coverage
- Uncertainty: Shrinkage-regularized Mahalanobis covariance with adaptive alpha = d/(d+n_k)
- ACT: Symmetric threshold updates with mean-reversion to prevent monotonic drift
- UP-UGF: Random projection LSH for O(N log N) approximate redundancy scoring
- Explainability: Historical penalty tracking replacing magic-number fallbacks
- Calibration: Bootstrap temperature estimation for cold-start autonomous operation
- Production: ONNX export, memory profiling, miniImageNet benchmarks, baseline comparisons
🏆 What Sets AdaptShot Apart¶
| Dimension | Conventional AI | AdaptShot |
|---|---|---|
| Hardware | GPU cluster required | CPU-only, <250MB RAM |
| Data per class | Thousands to millions | 5–50 images |
| Connectivity | Cloud-dependent | Fully offline |
| Confidence | Overconfident softmax | Calibrated ECE with conformal guarantees |
| Learning | Retrain from scratch | Continual via human corrections |
| Memory | Unbounded growth | Bounded via UP-UGF with LSH acceleration |
| Energy | Hidden carbon cost | Transparent per-inference Joule tracking |
| Trust | Black-box output | Conformal sets + uncertainty report + explanation |
| Cost | $10,000+ in compute | $0 (runs on existing hardware) |
| Quality Gates | Often absent | ruff=0, mypy strict=32 files clean, pytest=92 passed |
🔮 Roadmap¶
v0.2.0 (Current)¶
- [x] ONNX export for torch-free inference
- [x] Memory profiling with tracemalloc + psutil
- [x] miniImageNet benchmarks with baseline comparisons
- [x] Full production hardening of all algorithmic subsystems
- [ ] Swahili documentation translation
- [ ] Federated buffer sharing for multi-device deployments
v1.0.0 (Target: 2027)¶
- Field pilot results from 3+ NGOs in Tanzania, Kenya, and Uganda
- Peer-reviewed ablation studies
- Carbon-neutral CI/CD pipeline
- Community governance board
v2.0+ (Vision)¶
- Neuromorphic backend support
- Event-based vision for DVS cameras
- Multilingual, low-literacy UI extensions
- Integration with national healthcare/agriculture systems
👨💻 About the Creator¶
Johnson Christopher Hassan is a self-taught AI research engineer and diploma student in Computer Engineering at Mbeya University of Science and Technology, Tanzania. He built AdaptShot with a standard laptop, unreliable electricity, and determination—proving that world-class engineering doesn't require Silicon Valley resources.
- 📍 Mbeya, Tanzania 🇹🇿
- ✉️ johnson2006christopher@gmail.com
- 🐙 GitHub
🤝 Get Involved¶
- Use it: Deploy AdaptShot in your community. Share your results.
- Contribute: Submit a PR, write a tutorial, or translate documentation.
- Research it: Publish ablation studies, comparisons, or extensions.
- Teach it: Use AdaptShot in your classroom or workshop.