Behind PSR / The person behind the method
I build production AI systems for domains where a wrong output is expensive. Models are cheap. Taste, evals, and architecture are not.
Production AI and platform engineering across health tech, industrial control, and language-model systems. Experience at ISHI Health, Nophin, and Phaidra informs the questions PSR asks.
I’m Prabhjot, founder of Production Systems Review. PSR turns the questions I bring to production AI into a structured first-pass review.
The workspace produces automated analysis. A review by me is a separate conversation, grounded in the actual system and its evidence.
Talk to Prabhjot →Career evidence
Selected work from my published career and project record. These are engineering experiences, not PSR customer engagements or endorsements. Reported results belong to their specific evaluation or operating context.
Production LLM platform / Nophin
Built around explicit evaluation sets, targeted fine-tuning, and portable inference paths using Ray Serve and Modal.
A repeatable path from targeted fine-tuning and evaluation to production inference. Employer evaluation data is not published here.
Mission-critical industrial control / Phaidra
Designed real-time caching for model training, deployed reinforcement learning on ML accelerators, and instrumented the path with Grafana and Prometheus.
An instrumented data path for industrial model training and control, with monitoring in Grafana and Prometheus.
Applied RL research / AWS-funded ADAS-E
Combined bird's-eye sensor data, custom reward functions, tuned DQN variants, and distributed training across five GPUs.
The CARLA experiments exposed a reward-design trade-off: the agent prioritized speed over lane discipline. The write-up compares DQN, SAC, and PPO in simulation.
Read the project →Jan 2025 — Present
Leading the AI/ML platform architecture for a health-tech company, building intelligent systems that scale.
Jun 2024 — Nov 2024
Led AI infrastructure at a YC-backed startup. Designed cloud-agnostic ML platforms, fine-tuned LLMs to 100% accuracy on evaluation sets, and built model inference pipelines with Ray Serve and Modal.
Dec 2022 — Aug 2023
Led a team of 20 engineers building CRM systems for enterprise VoIP clients. Architected microservices for four different platforms and directed a full-stack training program.
Jan 2022 — Sep 2022
Built data pipelines handling 30GB daily for in-memory graph databases. Increased industrial operations efficiency by 20% using reinforcement learning on machine learning accelerators.
Oct 2020 — Dec 2021
Developed custom features from clickstream and transactional data for predictive analytics. Achieved a 30% increase in client conversions through data-driven statistical modeling.
Jun 2020 — Sep 2020
Investigated hard drive failures using machine learning. Built predictive maintenance models with Random Forest and Decision Trees, utilizing advanced sampling techniques.
Jan 2020 — Sep 2020
Created document representations using word2vec and multi-class classification with SVM. Built a dashboard for 100K documents with virtual scrolling and search.
Jan 2018 — Aug 2019
Led the Flyhomes.com search API and CRM products serving thousands of users. Managed mobile app development across iOS and Android with React Native.
Jan 2017 — Jan 2018
Built core platform features and data migration pipelines to AWS S3, optimizing storage and reducing man-hours significantly.
Jun 2015 — Jan 2017
Implemented NLP techniques using Stanford's CoreNLP for sentiment analysis for Fortune 500 companies. Built ETL pipelines processing thousands of PDFs daily.
May 2014 — Jul 2014
Research internship at France's premier paper technology institute. Worked on process optimization and materials science — the analytical foundation for a career in software and AI.
Career outcomes are self-reported. The underlying employer evaluation sets and operating records are not published here.
Published technical work
Stanford University / CS224R
Applied SFT, DPO, and GenRM to fine-tune Qwen 2.5 0.5B for mathematical reasoning. DPO achieved 45% win rate against the base Instruct model using H100 GPUs with FP8 training and Ray.io DDP.
Stanford University / CS224N
Enhanced small language model reasoning through teacher-trajectory SFT, cheatsheet summarization, and GRPO. Raised Qwen-7B accuracy on Game of 24 from 4% to 55% using distributed training on H100 GPUs.
Stanford University / CS231N
Predicted physiological brain age from MRI scans using ResNet transfer learning and custom 2D/3D CNNs. Achieved validation error as low as 1–4 years on 3,321 subjects from three medical datasets.
Stanford University / CS221
Solved the OpenAI Gym Lunar Lander environment using Deep Q-learning variants. Compared DQN, Double DQN, and Dueling DQN, achieving consistent 200+ scores — placing 2nd on the OpenAI Gym leaderboard.
University of Minnesota / Reinforcement Learning
Built an Automated Driving Assistance System for emergency vehicles using RL in the CARLA simulator. Trained DQN, SAC, and PPO agents to navigate high-speed emergency scenarios. Funded by AWS Research.
University of Minnesota / Deep Learning
Developed deep learning models to detect sleep and wake states from wristband accelerometer data. Implemented RNN/LSTM architectures with Gaussian kernel approaches using distributed computing. Funded by AWS Research.
Shorter working notes on models, evaluation, and architecture are published on X.
Working principles / In the founder’s words
Artificial Intelligence Graduate Certificate · 2025
CS224R Deep RL, CS224N NLP, CS221 AI Principles, CS231N Computer Vision
M.S. Computer Science, Mathematics Minor · 2024
Deep Learning, Reinforcement Learning, Big Data Engineering, Algorithms
B.Tech. Engineering · 2015
Work directly with the founder
To discuss a system with me, share the decision you’re facing and the kind of review you need. The automated workspace does not book or include a human review.