Foundations
Pre-training
Data, tokenization, model architecture, and the systems that create foundation models.
Explore on this site01 / Point of view
The interesting work begins where the demo ends.
I care about the full learning loop: the data we choose, the worlds we ask models to navigate, the failures we measure, and the systems that carry those lessons into production.
Research should leave the notebook.
— VK
02 / Research paths
Follow the model lifecycle.
Three connected bodies of work: how models learn, how they improve through feedback, and how they run reliably in production.
TrainRL
Post-training
Reinforcement learning, environments, graders, evaluations, and specialist agents.
Open trainrl.com ↗Serving
Inference
Paged KV caching, continuous batching, and fused kernels—how to keep the GPU busy at all times.
Explore on this site03 / Field notes
Ideas from the workbench
Notes on alignment, RL environments, and the signals that help language models improve through experience.
One machine. Two different jobs.
Prompt, environment, and grader form the shared architecture underneath evaluation and reinforcement learning.
Read issue 01 ↗Current research focus
What the environment makes observable
- 01 Planning across multiple steps
- 02 Choosing and using the right tools
- 03 Recovering when an action fails
- 04 Reaching a verifiable outcome
01 The shared structure
Measurement and learning belong in one conversation.
Evals tell researchers where a model is. RL environments create the experiences that move it. The nine-part series begins with the architecture they share, then follows the signal through graders, long-horizon agents, continual learning, human judgment, and trajectory design.
04 / Selected work
Building across the whole AI stack.
Research, data, infrastructure, deployment—the hard problems rarely respect team boundaries.
2025—Now
Senior AI/ML Engineer
Labelbox
Translating frontier-lab requirements into production RL data and environments; evaluating capability gaps across terminal, software, and cognitive tasks.
RL gymsEvaluationRLHF / RLAIF
2024—2025
Principal AI Developer
HTC Inc
Built hot-swap model adapters that cut deployment time by 80% and architected an end-to-end AI workflow management system.
Model adaptersFine-tuningMLOps
2023—2024
Senior Applied AI Scientist
Usable Machines
Built on the WhiteRabbitNeo foundation model and reduced task perplexity from 13 to 5 through preference optimization and task-specific fine-tuning.
Foundation modelsRLHF / RLAIFFine-tuning
2023
Adjunct Lecturer, NLP
Northwestern University
Created a build-an-LLM course and researched controlled legal-text generation with T5, reaching a ROUGE-L F1 score of 0.44.
NLPTeachingT5
2020—2021
Principal AI Developer
Code Techniq
Led an eight-person engineering team, built browser-based vision systems, and designed NLP pipelines for knowledge and medical applications.
BERTTensorFlow.jsML systems
05 / About
Researcher’s curiosity.
Engineer’s discipline.
I’m Vikram Kharvi, an applied AI researcher and engineer in California. I’ve worked across the AI lifecycle—from training and preference optimization to evaluation, deployment, and monitoring.
I hold a Master of Artificial Intelligence from Northwestern University and a Bachelor of Information Science Engineering from PES University.
EducationNorthwestern UniversityMS, Artificial Intelligence · 3.8/4
RecognitionHack2Innovate WinnerSamsung · NVIDIA · PWC · T-Hub
CertificationNVIDIA Agentic AICertified Professional
PYTORCH✦TRANSFORMERS✦RLHF✦RL ENVIRONMENTS✦AGENT EVALUATION✦KUBERNETES✦PYTHON✦MODEL EVALUATION✦
PYTORCH✦TRANSFORMERS✦RLHF✦RL ENVIRONMENTS✦AGENT EVALUATION✦KUBERNETES✦PYTHON✦MODEL EVALUATION✦