AI engineer · Founder and CTO, Craton Labs
Utkarsh Tripathi.
I build ML systems that learn from feedback, run safely and run fast.
Applied ML and deep learning in production, from real-time computer vision on edge hardware to recommenders and forecasting. Now going deep on RLHF and post-training, sandboxed environments for AI agents, and inference optimisation.
- YOLOv7 inference on a low-power ARM board
- 15+ FPS
- YOLOv7 inference on a low-power ARM board
- lower cost than the commercial ANPR systems it replaced
- 50%
- lower cost than the commercial ANPR systems it replaced
- sales forecast precision from Prophet models
- +15%
- sales forecast precision from Prophet models
- lower reporting latency from PySpark pipelines
- 40%
- lower reporting latency from PySpark pipelines
Focus
What I work on.
- 01 Current focus
RLHF and post-training
Reward models, preference optimisation with PPO, GRPO and DPO, and verifiable rewards that hold up against reward hacking.
- 02 Current focus
AI sandboxing
Isolated environments where coding agents and RL policies can run untrusted code, scored by verifiers they can't escape or game.
- 03 Shipped
Applied ML and deep learning
Models that run in production: YOLOv7 licence-plate recognition on edge devices, two-tower retrieval with DeepFM ranking, and time-series forecasting.
- 04 Shipped
AI optimisation
Making models fit the hardware: real-time detection at 15+ FPS on a low-power ARM board, and now GPU kernels, memory maths and efficient serving for large models.
Selected work
Things I've built.
- Applied ML 01
A real-time recommender for an online store
A lambda architecture built on Kafka, Flink and Airflow.
Read → - ML and RL 02
Tradr: a trading terminal in Rust
A text interface built with ratatui for live market data, backed by an ensemble of models, and why Rust suited a data-heavy tool.
Read → - ML fundamentals 03
Machine learning algorithms in plain NumPy
Gradient descent through to gradient boosting, implemented without ML libraries.
Read → - Edge systems 04
Getting diagnostic data off trains on the Konkan line
LTE drops out for hours along the 740 km route. A unit built on an ESP32 and a Raspberry Pi Pico, under $30 in parts, stores telemetry and unloads it at each station the train passes.
Read → - AI product 05
StudyGrove: a study planner with a farm that dies if you skip
An AI writes daily plans from your own notes, a tamagotchi-style grove keeps you coming back, and short checks confirm you learned the material.
Read → - LLM evaluation 06
How I check whether an LLM's claims are backed by its sources
Splitting answers into single claims, testing each against the cited text, and probing for instructions hidden inside retrieved documents.
Read →
Experience
Where I've done it.
Five years across rail engineering, computer vision on the edge, insurance data and now my own company.
Full resume (PDF) →- Oct 2025 – now
Founder and CTO, Craton Labs
Building AI systems for banks, insurers and fintechs.
- Jan – Sep 2025
Data Engineer, RiskPe
PySpark pipelines over 10K+ daily policy and lead records cut reporting latency by 40%. Prophet forecasting lifted sales forecast precision by 15%.
- Jul – Dec 2024
Computer Vision Engineer, Habitat
Put YOLOv7 licence-plate recognition on BeagleY-AI edge boards at 15+ FPS, at half the cost of the commercial systems it replaced. Raised accuracy in poor light by 20% with data augmentation and tuned post-processing.
- 2021 – 2023
Systems Engineer, Siemens
Condition-monitoring analytics and failure prediction for rolling stock on the Vienna Metro, Sydney Airport Line and Indian Railways projects.
Toolkit
What I build with.
- ML and deep learning
- YOLOv7OpenCVLSTM and Transformer modelsXGBoostProphetTwo-tower retrievalDeepFMNumPy
- LLMs
- LangChainSentence-TransformersQdrantMilvusRAGASFastAPI
- Data
- PythonSQLSparkKafkaFlinkAirflowdbt
- Infrastructure
- AWSDockerKubernetesTerraformCI/CDRustESP32 and Pico
Contact
Working on RL, model training or inference?
I take on independent consulting work, and I'm glad to talk shop about RLHF, sandboxing and making models faster.