Hi, I'm
Pranav Chetan
Software engineer · Robotics + applied AI · Open to roles
I build AI that leaves the screen and does real work in the physical world: a robot that maps a warehouse in 3D and inspects its welds with no one at the controls, a 3D-scanning workstation built on a laser scanner I reverse-engineered, an AI analyst that answers questions about Congress without the data ever leaving the machine, and an F1 model that predicts the full finishing order before the race starts.
-
01 · Robotics
AMR
An autonomous robot that maps an entire warehouse floor in 3D, then drives itself zone to zone to inspect welds, with no one at the controls.
Reaches every goal in 96.5% of test runs, and 94.5% of runs finish without touching a thing.
More
AMR explores a warehouse on its own and maps the entire floor in 3D. Then it works through the floor zone by zone: it drives itself to every area with welds to inspect, scans each weld with a gimbal camera whose liquid lens refocuses in milliseconds, and sends the images back. No one at the controls.
I engineer its autonomy at every layer, from 3D mapping and route planning to the AI that drives it and the mission it runs. Try the console: watch it run an inspection, send it a goal, or drop an obstacle in its path.
MissionsWeld inspection · 3D floor mapping · drive-to-spot scans · box inspection · inventory counts · aisle patrols
Goal success96.5%Test runs where the robot reached every goal it was given
Obstacle avoidance94.5%Test runs finished without touching a single obstacle
200 test runs in a cluttered room with partitions and alcoves
- Python
- PyTorch
- Reinforcement learning
- ROS 2
- FAST-LIO2 SLAM
- Jetson Orin
- LiDAR
- RealSense
- YOLO11
-
02 · Robotics
UR5e inspection cell
Three cameras inspect the top and both sides of every box on a conveyor, and a UR5e robot arm removes the defective ones into a reject bin.
A three-stage AI pipeline (EfficientAD → YOLO → Qwen3-VL) decides, so no one has to watch the belt.
More
Three cameras on one arch inspect the top and both sides of every box on a conveyor. A three-stage AI pipeline decides whether a box is defective, and the UR5e arm removes the bad ones into a reject bin, with no one watching the belt.
3Cameras: top and both sides3AI stages: EfficientAD, YOLO, Qwen3-VL -
03 · 3D scanning
CloudClean
A 3D-scanning workstation that turns raw laser scans into measured, inspection-ready parts. Import a scan and Autopilot does the rest.
Reverse-engineered the scanner so it runs on Linux, then cut its scan error about 10×.
Pick a part to scan: click one on the table (the blue rings) or here
More
A workstation that turns raw scans from a metrology-grade laser scanner into cleaned, measured, inspection-ready parts, from any web browser, with an AI agent that can run the entire app for you.
AutopilotYou: import a scan
- Import
- Clean
- Merge
- Mesh
- Inspect
- Export
You: 1 stepAutopilot: the other 5, no clicks
Import a point cloud and that's the whole job. Autopilot cleans it, checks and merges multiple scans, builds the mesh, inspects the part against its CAD model if you've added one, and exports it, without you clicking a single button. In most scanning software, each of those is a separate tool you run and tune by hand.
The scanner's maker doesn't support Linux, so I decoded how it talks over USB and wrote my own driver from scratch.
~10×Lower scan error than my first driver89 fpsLive capture on Linux~28App tools the AI assistant can use~300Automated tests- Python
- Open3D
- OpenCV
- FastAPI
- React
- TypeScript
- three.js
- vLLM
-
04 · GM sponsored
Policy analytics
An AI analyst that answers plain-English questions about bills, lobbying and campaign finance, running entirely on one machine.
Turned minutes or hours of SQL and BI work into answers in seconds, each traced to its source.
More
A conversational AI analyst for congressional data, sponsored by General Motors. Policy analysts ask about bills, lobbying and campaign finance in plain English and get answers drawn from curated data, with no SQL or BI tools involved.
Everything runs on one machine: the language model, the search and the data. No question or answer is ever sent to the cloud.
100%Answers grounded in Gold tablesMin → secQuestion turnaround- Python
- RAG
- LLM
- SQL
- Medallion architecture
- Audit logging
-
05 · Machine learning
F1 prediction
A learning-to-rank model that predicts the full finishing order of a Formula 1 race from decades of results.
Scored 10–25% better on ranking metrics than accuracy-only baselines.
You drive · 2 laps against the 22-car grid
Replay · Monaco 2024, Leclerc's home win · real timing data
More
A race-outcome and driver-ranking system trained on decades of Formula 1 results, and judged the way search engines judge rankings, with NDCG and MAP, not raw accuracy.
10+Seasons of race data30–60Features per race entry3Model families compared+10–25%NDCG / MAP lift over baseline- Python
- scikit-learn
- Pandas
- Random forest
- Gradient boosting
- DNN
-
06 · About me
-
07 · Ask Pranav
Type a question on the monitor: projects, tools, what I am looking for. The terminal answers from this site's own pages.
About me
Who's building this
I take on problems where the model is only half the job. The other half is the data it stands on, the metric that proves it works, and the service that puts it in front of people.
I studied computer science at Wayne State University on the GM-sponsored project track. Today I'm an AI/ML engineer intern at L&T Technology Services, putting AI on autonomous robots, industrial arms and edge devices. I've built an inspection robot's full autonomy stack, a 3D-scanning workstation, ranking models and retrieval pipelines, and I don't call a system finished until I can prove it works.
- Education
- B.S. Computer ScienceWayne State University · Class of 2025
- Focus
- Robotics, applied AI, backend services, Linux systems
- Certifications
- Machine Learning SpecializationStanford & DeepLearning.AI · 2024
- Looking for
- Full-time roles in software, robotics and applied AI
Contact
Hiring? Let's talk.
I'm looking for full-time engineering roles in software, robotics and applied AI. Email is the fastest way to reach me.
pranch555@gmail.com