Engineer, Researcher, Image-Maker. — Anant Saraf
Brown | RISD Dual Degree: CS + Photography
I love digital images! And I love making software that deals with them, whether to make images from scratch (Graphics), make them more beautiful (computational photography) or to extract meaningful information from them (computer vision). This technical portfolio contains selected projects across research, coursework, and personal work.
cd to reach my fine art portfolio.
Lossy
Full Stack Image Sharing Platform · April 2026 · App development
A social-network app where, every time your image gets more viewers, it degrades in quality. Every view applies an incremental degradation step (JPEG recompression, CLAHE contrast shifts, noise injection, pixel sorting, color-space corruption, false-color overlays, eventually full-on pixel-swap chaos) so posts visibly decay over time. Join Lossy here.
- Designed and shipped a deployed full-stack social platform where image fidelity progressively decays with view count, exploring degradation as a social mechanic.
- Architected a server-side image degradation pipeline (NumPy, OpenCV, Pillow) with checkpoint caching and bounded replay.
- Built on a React/TypeScript frontend, async Python/FastAPI backend with PostgreSQL, and Cloudflare R2 + Workers for image storage and signed-URL delivery; packaged for Android via Capacitor.
Comp Photo Coursework
Computational Photography · Fall 2025 · Image processing
Coursework in computational imaging — coarse-to-fine Lucas–Kanade optical flow, HDR radiance merging with saturation-aware weighting, and planar/cylindrical panorama stitching with RANSAC-robust alignment.
- Panorama stitching estimates translation, affine, and homography transforms with least-squares and RANSAC, including EXIF-based focal length conversion for cylindrical projection.
- HDR tone-mapping preserves highlight and shadow detail on standard displays via saturation-aware weighting of the radiance map.
- Other assignments included Poisson blending, pyramid-based RGB image alignment, weighted-bilateral texture filtering, color transfer, and RAW image processing.
Ocean Eddy Detection Pipeline
UTRA: Geospatial Object Detection · Summer 2025 · End-to-end pipeline and model training
End-to-end SAR computer vision pipeline to detect ocean eddies in open water and marginal ice zones — packaged as a reproducible, config-driven workflow usable by non-technical domain experts. All a user has to do is enter min and max lat/lon coordinates, and the pipeline will download images, preprocess, run inference and postprocess to give the user ocean eddy sizes and locations.
- GeoTIFF ingestion via API, SAR-specific preprocessing, YOLOv11 inference with SAHI sliding-window tiling, and geospatial post-processing produce analysis-ready outputs.
- Fine-tuned on a 3,000-image dataset (including null samples) to ~90% precision, 78% recall, and 83% mAP@0.5 on a held-out test set.
- Partnered with 2 domain experts to translate scientific objectives into dataset and modeling decisions, and built curation tools to support iteration and handoff.
Ray Tracer
Graphics · Fall 2025 · Rendering pipeline
C++ ray tracer built from scratch — designed the renderer architecture and implemented ray generation, intersections, shading, and recursive reflections.
- Debugged floating-point precision and reflection-recursion failures using scene-based visualization to improve numerical robustness across complex reflective scenes.
- Anti-aliasing uses mipmapped texture filtering with cached mip levels and separable kernels to improve sampling efficiency.
- Supersampling strategy (grid, random, stratified) selectable per scene.
Real-Time Graphics Renderer
Graphics · Fall 2025 · Rendering pipeline, shading effects, performance optimization
End-to-end real-time OpenGL pipeline — GPU instancing, shading, post-processing, and a custom LUT — merged with three collaborators' subsystems during an integration phase.
- GPU instanced rendering reduced per-frame time from 800 ms to 0.015 ms, enabling real-time rendering of 153,000 cylinders.
- Post-processing pipeline: color grading with a custom-designed LUT, toon (quantization + Sobel) filtering, and multiple screen-space passes.
- Benchmarked scalability by ramping instance counts and scene complexity.
Gaussian Splatting Pipeline
Computer Vision · Spring 2025 · Training pipeline and differentiable renderer
3D Gaussian Splatting pipeline in PyTorch for differentiable 3D reconstruction — defined model structure, initialized parameters, and optimized end-to-end.
- Mixed-precision training on an HPC cluster, monitored across 58 hours with checkpointing.
- Diagnosed and debugged numerical stability issues across differentiable rendering stages (projection, covariance, alpha blending).
- Ran experiments integrating SSIM-based perceptual loss alongside the standard reconstruction objective.
Lung Segmentation + Disease Classification
JISIASR: Medical Imaging · Summer 2024 · Model research and training experiments
Repeatable training and evaluation pipeline for chest-X-ray lung segmentation, working under small-data and label-quality constraints. Benchmarked U-Net variants against a ResNet-50 classifier baseline.
- Ran controlled experiments on preprocessing (CLAHE) and training choices (augmentations, early stopping) to reduce overfitting and improve validation behavior; compared U-Net, ResUNet, and TransResUNet variants and summarized results for handoff.
- Applied Integrated Gradients and LIME on the ResNet-50 classifier to probe model sensitivity and failure cases on bacterial vs. viral pneumonia.
Intro to Software Engineering (TA)
Teaching Assistant: CSCI 0320 · Fall 2025 · Design labs, mentor 100+ students, evaluate sprint demos
Course material covers API development, socially responsible computing, software architecture, data validation, version control workflows, and agentic AI-assisted programming.
- Ran weekly labs and office hours for 100+ students, mentoring teams and reviewing sprint demos.
- Gave structured feedback on testing, code quality, Git workflow, and system design.
Brown Course Review Webapp
Intro to Software Engineering · Fall 2024 · Backend API and testing
Java (Spark) HTTP API + Firebase data layer serving reviews for 2,700+ courses and 1,600+ professors.
- Unit and integration tests at 75–100% coverage server-side, plus Playwright end-to-end coverage of critical flows.
- Collaborated with a team of 4 on a concurrency-safe, user-centric app, optimizing API queries and retrieval across sprint demos.