About me
Hi! I’m Pranav Balaji and I’m interested in applied ML research and engineering. I’ve worked mostly in computer vision on data-efficient and generalizable training approaches. Apart from research, I’ve also built 0-to-1 AI products and scaled them.
In my undergraduate thesis at Inria, supervised by Dr. Antitza Dantcheva and Dr. Abhijit Das, I researched generalizable deepfake detection using contrastive learning and self-supervised auxiliary training on video transformers.
Then, I co-founded a startup on talking head generation. As the CTO, I was involved in researching latent diffusion models, rendering models and creating a dataset. I also architected the deployment infrastructure to run multiple deep models at minimal latency and cost, along with some MVPs on ultra-low-bandwidth video calling and education module creation.
Now at Space42, I work on computer vision and LLM agents for satellite imagery. This includes an active learning system for data-efficient (or rather, human-effort-efficient) training with adaptive tiling and uncertainty sampling. I also work on the engineering side - deploying scalable inference and training systems.
So far in my career, I’ve experienced a wide range of the stack - from researching model architectures to building products out of them to scaling them. Now that I’ve covered the breadth, I think it’s time for depth, wherever the most interesting problems are. More often than not, I’ve found them to be in research.
My hobbies on the other hand are all breadth. I do origami and speedcubing, play the violin, keep small pets (fish, snails, mantids), work on personal projects (personal productivity app, quant trading experiments, etc.), travel and go scuba diving - I have the Advanced Open Water Diver certification. I’ve also been learning Japanese intermittently since high school and have a deep fascination with the language.
- Computer Vision
- Data-Efficient Learning
- Scalable ML
- LLM Agents
Bachelor's in Computer Science and Engineering, 2019 - 2023
BITS Pilani
Extra Interests
Language
PADI Advanced Open water certified
The art of paper folding
Mostly the 3x3
Experience
- Re-architected the geospatial inference engine with Ray (tiling, raster I/O, model-specific processing): 7–10× faster on the same hardware, 4× fewer components, higher observability and 97% fewer user error reports over 1 year.
- Developed a geospatial active learning platform for data-efficient training with adaptive tiling and uncertainty sampling. Cut down annotation and training cycles from days to hours.
- Built a general-purpose, agent-first geospatial platform for acquisition, raster and vector intelligence, with an LLM agent to run custom workflows.
Deeptech GenAI startup on facial video synthesis, incubated by Inria Startup Studio.
- Involved in training latent diffusion and rendering models, and evaluating pretrained models across stages of the synthesis pipeline. Also collected a new video dataset to train higher resolution models.
- Built web and ML infrastructure to run 5+ deep models with sub-minute latency and minimal cost per request.
- Built MVPs for ML-encoded video calling at 1000× lower data rates than standard video streaming, and an AI education module creator turning raw content into teachable presentations in minutes using Generative AI.
- Undertook 20 days of Inria Startup Studio’s camps covering market analysis, pitching and fundraising.
Thesis on deepfake detection that generalizes to unseen manipulations, supervised by Dr. Abhijit Das (BITS Pilani) and Dr. Antitza Dantcheva (Inria). Led to a first-author ICCVW 2023 paper and the video experiments in a WACV 2024 paper.
- Generalizable deepfake detection using multi-task and contrastive learning with NCE and cross-entropy losses.
- Applied self-supervised auxiliary training with VideoMAE for deepfake detection, achieving stronger zero-shot transfer than the DFDC winner.
Teaching assistant for the course Natural Language Processing.
- Proposed Transformers coverage and helped prepare lecture notes.
- Conducted a 1-hour PyTorch demo: Code
AI for Social Good.
- Predicted treatment-dropout risk for the Indian government’s national Tuberculosis Adherence program, to most efficiently allocate human follow-ups on patients.
- Optimized models for a custom metric on imbalanced data using surrogate loss functions.
Accomplishments
A 3 day ML Hackathon conducted by Amazon with 26000+ participants from top universities. The task was to predict the product length in mm given the product title, description, and bullet points as text.
- Leaderboard: Team Fishes
- GitHub Repository: AmazonML
A nationwide Hackathon conducted by Amazon with 2000+ participants from top universities. Took part under the theme Covid-19 and Technology and won. Pitch:
- An online game-ified Virtual Fitness application designed for lockdown times
- Real-time Machine Learning on the webcam feed for posture identification
- Video call system with live ranking against a friend
- GitHub Repository and related links: vFit
