Deep-tech startup
★★★★★ 5.0
Quantum-inspired optimization
Validated the math and code of a novel optimizer and shaped it into an investor-ready product.
Applied AI studio · Now taking projects
From research
to real-world AI.
We build AI that sees, senses and acts: vision, robotics, audio and sensor intelligence, and state-of-the-art R&D. Shipped end to end, from cloud to device.
- 100%Job Success on Upwork
- 5.0★on every reviewed job
- 3industry AI roles
- 4live in-browser demos
01 — Track record
Client work, rated 5.0.
Delivered through Upwork for startups and hardware companies.
ML tooling product · Private
★★★★★ 5.0
Training-efficiency research
Controlled A/B experiments proving where a learning-rate scheduling product speeds up training.
Long-term R&D partner · Private
★★★★★ 5.0
Long-term deep learning partnership
Ongoing model development for our longest-running client. Details are private.
Hardware company
Delivered on scope
Real-time vision for a dashcam
Real-time detection and video processing for an in-vehicle camera product.
Verified on our Upwork profile
02 — How we work
How a project runs.
-
01
Feasibility
We review your data and goals and confirm what is achievable before you commit.
-
02
Prototype
A working model on your data, with clear metrics.
-
03
Harden
Edge cases, optimization and accuracy checks.
-
04
Deploy & hand off
Running on your stack, documented for your team.
03 — What we do
Applied AI, end to end.
Systems that see, sense and act, designed for production from day one.
Edge AI deployment
Quantized, optimized models running in real time on low-cost hardware.
- INT8
- ONNX
- TensorRT
- Vendor NPU toolchains
Perception & vision systems
Detection, tracking and scene understanding for vehicles, cameras and robots.
- YOLO
- ByteTrack
- Landmarks
- Segmentation
Visual inspection & anomaly detection
Defect detection from photos of good parts only, at line speed on a CPU.
- PatchCore
- Label-free
- CPU-ready
R&D: paper to product
State-of-the-art methods rebuilt from scratch, verified, and turned into products.
- Re-implementation
- Ablations
- Benchmarks
Robotics & autonomous agents
Reinforcement learning agents and simulation environments that learn by doing.
- Deep RL
- Custom Gym envs
- Sim
Signal & audio intelligence
Models on raw audio and sensor streams: voice-fraud and anomaly detection.
- Raw audio
- Graph attention
- Time series
04 — Featured production work
In-vehicle safety perception, fully offline.
A confidential production system, shown by architecture only.
The problem
Catch fatigue, distraction and safety violations in real time, with no network and no video leaving the vehicle.
What we built
Parallel vision models on one low-power chip, turning frames into graded safety events with on-device video clips and logs.
Engineering highlights
- Accuracy-preserving compression: a calibrated low-precision pipeline that keeps on-device accuracy aligned with the full-precision model
- Multi-model real-time scheduling: several concurrent networks sharing one low-power accelerator at stable frame rates
- Train-to-device parity: numerical verification across the whole pipeline, so deployed outputs match what was validated in training
- Temporal event reasoning: sequence modeling that turns noisy per-frame signals into reliable, severity-graded alerts
- Offline-first reliability: self-contained capture, logging and evidence retention with zero cloud dependency
Per release
TrainPyTorch
ExportONNX
QuantizeINT8 + calibration
On device · no network
Camerain-cabin stream
Face & landmarks
Device detection
Restraint check
Temporal modelConv1D over frames
Event logicseverity tiers
Incident clip + logpre/post buffers
05 — Lab work
Products we have built. Try them live.
Each one was built in-house, and the demos run right in your browser.
Autonomous agents · Automated game testing
DriftMind
An AI player that teaches itself any game just by watching the screen, with no rules and no game-specific code. Point it at a new build and it plays thousands of sessions overnight to surface bugs, exploits and difficulty spikes before your players do.
- 100%
- success in 3D navigation
- 18+
- targets per run in a shooter
- 43%
- never-seen levels cleared
Private build · demo on request
AI quality inspection
DefectScope
An automated inspector for production lines that learns from photos of good parts only, then pinpoints scratches and dents on anything that doesn't match. No defect library and no labelling, and it outperforms Intel's open-source reference.
- 94%
- defects caught
- 0
- defect photos needed
- 50 ms
- per part, no GPU
Deepfake voice detection
VoiceShield
Tells a real voice from an AI clone in seconds: a security layer for call centres, voice logins and identity checks. It catches cloning methods it has never seen before.
- 1.4%
- error rate
- 11
- unseen cloning methods
- 0.14%
- false alarms on real voices
Game graphics enhancement
PixelForge Engine
AI remastering for game graphics: old, blurry textures come back as crisp 4× HD in a fraction of a second each. Repeating tiles stay seamless, edges stay clean, and it outperforms the industry-standard upscaler.
- 4×
- sharper resolution
- 15×
- lighter model
- ~9×
- faster
Satellite intelligence
MapWise
Turns raw satellite images into land-use maps, measuring farmland, forest, water and urban growth for agriculture, insurance and environmental work. Runs entirely in the browser, so imagery never leaves the device.
- 6
- land types mapped
- 85.9%
- pixel accuracy
- 0
- server uploads
More research & projects 5
Sensor networks
Conformal Forensics
Attack-resistant anomaly detection for sensor networks, with guaranteed limits on false alarms.
Training efficiency
Adaptive LR Engine
An engine that adapts how models learn as they train, benchmarked against standard methods across frameworks.
Research · in progress
Multiscale anomaly detection
A new architecture that spots anomalies at every scale, developed with a university professor for a top computer vision conference.
Sports vision · in progress
Ball tracking from synthetic data
Follows a fast-moving ball and hears every hit. Trained on simulated courts, so no manual labelling is needed.
Earlier work
Generative AI, agents & more
Image generation, game-playing agents, an object detector built from scratch, and a deployed medical-fraud detector (94% accuracy).
All repositories ↗
Priyanshu Singh
Founder & Lead Engineer
- AvailabilityRemote · works across time zones
- EducationBachelor's in AI / ML
- IndustryComputer vision, AI research and generative AI roles
- Also5+ years of 3D & Blender, now used for synthetic training data
06 — The studio
Research depth, production focus.
M2M stands for Model2Metal: taking AI from the research model all the way to the system it runs on. We build models from the ground up, validate them on real benchmarks, and deploy them on cloud or embedded hardware.
You work directly with the engineer building your system.
PyTorchONNXTensorRTINT8 quantizationEmbedded NPUsYOLOOpenCVtimmUltralyticsONNX Runtime WebDockerAWSBlender
07 — Start a project
Let’s talk about your project.
Share a few details about what you're building and we'll get back to you within a day.
Priyanshu SinghFounder & Lead Engineer. Every enquiry comes straight to me.
Prefer email? singhpriyaanshu308@gmail.com