---
title: Slipstream
emoji: 📈
colorFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: 6.16.0
app_file: app.py
pinned: false
license: cc-by-4.0
tags:
  - track:backyard
  - sponsor:openbmb
  - sponsor:nvidia
  - sponsor:modal
  - achievement:welltuned
  - achievement:offbrand
  - achievement:sharing
  - achievement:fieldnotes
---

# Slipstream

This space provides an interactive presentation and live demo of the Slipstream project-controls forecasting benchmark: a new benchmark for forecasting a project's final cost (EAC) and finish period from mid-flight Earned Value data, an agentic layer that reconciles the forecasting tools, and the distillation of that agent into small models for edge, air-gapped, and on-device forecasting.

The application is built with a buildless Preact frontend served by a Gradio Server (`gr.Server`), which also exposes the benchmark results through typed API endpoints. The headline edge agent is `openbmb/MiniCPM5-1B` distilled into a project-controls forecasting agent ([`slipstream-minicpm5-1b-evm`](https://huggingface.co/build-small-hackathon/slipstream-minicpm5-1b-evm)). Off the shelf, the 1B base model returns a usable forecast less than 2% of the time; after distillation, it reaches ~99% validity with a median cost error of roughly 2.7%, matching the traditional Earned Schedule baseline. It serves as the default model in the live demo. We also distilled `nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16`, a hybrid Mamba-2 + attention model, into [`slipstream-nemotron3-nano-4b-evm`](https://huggingface.co/build-small-hackathon/slipstream-nemotron3-nano-4b-evm) - our strongest student model, achieving about 2.37% cost error and a 0.61-period finish error. It was trained on its dedicated CUDA stack on Modal and is selectable in the live demo.

DEMO: https://youtu.be/v-k7lxXskTA

## Try it live

The final slide runs a real held-out project through the agentic layer live on Modal, next to Earned
Schedule, TimesFM and TabPFN, and compares every forecast against the true outcome. Each run
cold-starts a GPU, so expect roughly 5-7 minutes; the methods stream in as they finish.

## Social post

Hackathon social post: https://x.com/NZXW63TF/status/2066647669540360315

Disclaimer: I do not really use social media. This X / Twitter account exists only so I can scrape
tweets from lists.

## Links

- Write-up / article: https://huggingface.co/blog/build-small-hackathon/slipstream
- Distillation dataset: https://huggingface.co/datasets/build-small-hackathon/slipstream-evm-sft
- MiniCPM5-1B agent: https://huggingface.co/build-small-hackathon/slipstream-minicpm5-1b-evm
- Nemotron-3-Nano 4B agent: https://huggingface.co/build-small-hackathon/slipstream-nemotron3-nano-4b-evm
- Gemma-E2B agent: https://huggingface.co/build-small-hackathon/slipstream-gemma4-e2b-evm
- Organisation: https://huggingface.co/build-small-hackathon

The dataset is CC-BY-4.0; each model inherits its base model's licence. Everything was trained,
evaluated and benchmarked on Modal including the live demo in this space.

## Run locally

```bash
pip install -r requirements.txt
python app.py     # http://localhost:7860
```

The live-demo backend is a separate deployed Modal app (`pipeline/agent/demo_modal.py`); set
`SLIPSTREAM_DEMO_URL` to point the Space at your own deployment if you redeploy it.