I take open-weight vision and multimodal models from raw data to a production endpoint on infrastructure you control. Your data stays inside, and the code and weights stay yours.
Trusted by teams at
Every request ships your images and documents to someone else's servers, under someone else's terms.
Per-token pricing punishes you for using the product more. Self-hosted inference costs what the hardware costs.
You cannot fine-tune it, pin a version, or take it with you when the vendor changes the pricing.
Representations learned from your raw images, no labels required. Self-supervised training in the I-JEPA line, and backbones like DINOv3 carried into domains they were never trained on, multi-spectral among them.
See the I-JEPA buildOpen weights adapted to your data. LoRA and QLoRA on vision-language models, segmentation and detection, from satellite imagery to surround-camera video, on the VRAM you actually have.
See it in the workYour model, your machine, your URL. An OpenAI-compatible endpoint on your own hardware, vLLM picked automatically on NVIDIA GPUs, plus streaming pipelines that keep long video inside a real memory budget.
See SursumAIWe start from the constraints that decide everything else: your data, your hardware, your latency budget and what can never leave your network.
The right model is picked and trained on your data, pretrained from scratch or finetuned from open weights, then measured against the baseline on your own examples.
It goes live behind an endpoint you control, running on your machines, with the code and weights handed over to you.
A focused read on whether open weights win for your case.
A real model, serving behind an endpoint on your machines.
Hardened, documented and handed over to your team.
Teams that need a vision model working now, without building an ML department first.
Where the data cannot leave the building, and an external API is not an option.
Multi-spectral segmentation, cloud removal and land-cover analysis at scale.
Bird's-eye-view perception, real-time depth and reconstruction from video.
Your model. Your machine. Your URL. Deploy open-weight and custom models on your own hardware and get an OpenAI-compatible endpoint. No terminal, Docker, or GPU knowledge needed. Picks vLLM automatically on NVIDIA GPUs.
Visit sursumai.pages.devStreaming pipeline that lets Depth Anything 3 run on long video sequences and large-scale scenes under tight CPU/GPU memory budgets, chunking frames and carrying state across chunks for stable online inference.
View on GitHubTwo 3D reconstruction pipelines compared on the same multi-view capture: SAM3 segmentation feeding COLMAP sparse and dense MVS, against 3D Gaussian Splatting converted to mesh with SuGaR. Both end in textured geometry.
View on GitHub
Stacks cloudy Planet imagery over clean Sentinel-2 of the same place. Detects the clouds, opens a hole through them and the surrounding haze, matches brightness by histogram, and composites with a feathered edge so the seam disappears.
View on GitHubBird's-eye-view perception system replicating Tesla FSD's surround-camera fusion pipeline for real-time vehicle and lane reconstruction.
View on GitHub
Multi-spectral semantic segmentation of satellite imagery for Earth Observation, built for Privhti EO's geospatial intelligence platform.
View on GitHubLocal video summarizer that runs entirely on-device. No data leaves your machine. Extracts key moments and generates structured summaries from any video file.
View on GitHub
Adapting DINOv3 self-supervised features to multi-spectral imagery for dense similarity mapping, enabling zero-shot land-cover analysis without labeled data.
View on GitHub
CLI for fine-tuning and fast inference of Qwen3-VL with LoRA/QLoRA, adapting open-weight vision-language models to domain data with minimal VRAM.
View on GitHub
Deep Learning Engineer | Pretraining, Finetuning, Inference & Serving Vision Models
Brasil · 4,590 followers
No. Smaller models run on CPU through llama-server. If you have an NVIDIA GPU, SursumAI picks vLLM automatically for far higher throughput.
No. Training and inference run on your machines. I work inside your environment and nothing is uploaded to a third party.
You do. Everything built in the engagement is handed over at the end, including training code, serving code and model weights.
If a closed API covers your case and the data is not sensitive, use it. Open weights win when data cannot leave, when usage is heavy enough that per-token pricing hurts, or when you need to fine-tune and own the model.
Vision and multimodal models: Qwen-VL, DINOv3, Depth Anything, SAM, Prithvi, YOLO, and whatever fits the problem.
Book a 30 minute call. We go through your data, hardware and constraints, and I tell you whether open weights make sense for your case.
Have a model that should be in production by now?
