Deep Karkhanis

AI Researcher

San Francisco, California, United States5 yrs 2 mos experience

Most Likely To SwitchAI ML Practitioner

Key Highlights

Core contributor to Gemini AI models at Google DeepMind.
Developed foundational LLMs at Abacus.AI.
Achieved >98% accuracy in face recognition algorithms.

Stackforce AI infers this person is a Machine Learning and AI expert with a focus on Large Language Models.

Contact

deepkarkhanis@gmail.com +917208121165 LinkedIn

Skills

Core Skills

Large Language Models (llm)Machine LearningReinforcement Learning

Other Skills

AlgorithmsAndroid StudioArtificial Intelligence (AI)Artificial Neural NetworksCC++Computer ScienceComputer VisionConvolutional Neural Networks (CNN)Deep LearningDigital Image ProcessingFormal MethodsFormal VerificationImage ProcessingJava

Experience

Google deepmind

Research Scientist

Jan 2024 – Present · 2 yrs 2 mos · Mountain View, California, United States

Gemini Team
Core Contributor to Gemini 2.5, 2.0, Gemma 3
End to end Gemini Training (Pretraining, SFT, RL, Thinking)
Gemini Code Generation
Jules: The AI Software Engineer (https://jules.google/)
Taking Gemini to SWE-Bench No. 1
RL research:
Pass@k Policy Optimization [Neurips Spotlight '25]
DPO-Positive [ICLR '25]

Large Language Models (LLM)ResearchMachine LearningAlgorithmsDeep LearningArtificial Intelligence (AI)+1

Abacus.ai

Machine Learning Researcher

Jan 2022 – Jan 2024 · 2 yrs · San Francisco Bay Area

Working on the GenAI team. Building Open Source Foundational LLMs
Key LLMs released: Smaug-72B, Giraffe-32k

Large Language Models (LLM)Machine Learning

Tower research capital

Quantitative Trader

Jan 2022 – Jan 2022 · 0 mo · New York, New York, United States

Microsoft

Microsoft AI Researcher

Jan 2020 – Jan 2021 · 1 yr · Bengaluru, Karnataka, India

Worked under Dr. Harsha Vardhan Simhadri and Dr. Ravishankar Krishnaswamy as a Machine Learning Engineer and designed a system for >98% accuracy kNN search on datasets with >106 points & >100 dimensions with high rate of inserts and deletes
Employed a hybrid model which performs incremental Vamana Indexing (Microsoft DiskANN) on top of streaming FAISS clustering
Improving the system to provide high QPS on low (<5MB) memory servers suffering from high pre-emption rates and limited disk access (no SSDs)
The expected number of distance calculations are 50% less than the current state of the art low memory solutions

Kwikpic.in

Jan 2019 – Jan 2020 · 1 yr

Adapted leading face-recognition algorithms to work on Indian faces
Handled varied lighting conditions and achieved an accuracy of >98% for indoor & >95% for night events
Software engineer , Machine learning engineer

Rwth aachen university

Scientific Researcher

Jan 2019 – Jan 2019 · 0 mo · Aachen, North Rhine-Westphalia

Combined concepts of Stochastic Model Checking & Counting-SAT to compute Bounded Reachability Probabilities in MDPs
Achieved 10x faster solving by designing a succinct CNF encoding for Markov Chains using the transition probabilities BDD

Institute of science and technology austria

Scientific Researcher

Jan 2018 – Jan 2018 · 0 mo · Austria

Improved the POMCP algorithm to create an online MEMDP solver and established its superiority over POMDP solvers
Exploited the sparse transitions in MEMDPs to have faster belief updates
[O(n) as opposed to O(n2 )]
Solver was 50x faster & a 20x better environment detector with higher success & crash-less navigation on Hallway benchmarks
Published findings at ICAPS 2020