Dry Run

Test your plans before you bet the shop on them.

Describe your business in one line. Tavily finds the news that hits it, NVIDIA Nemotron (served by Nebius Token Factory) proposes plans, and a deterministic simulator tests every plan in several futures. You get a decision memo with its sources and an honest list of what it doesn't model.

Live demo: https://dryrun-1hny.onrender.com · Built for the Nebius x NVIDIA Global AI Hackathon (Best Apps and Agents) · MIT license

For judges: try it in one minute

1. Open https://dryrun-1hny.onrender.com. The free host can take about 50 seconds to wake up.
2. Pick an example, such as Coffee roaster in Austin, Texas. It replays a run recorded live with Nemotron, so it is free and always works.
3. Press "Check live news" to make a real Tavily call. Scroll the landing page to "We tested it on the past" for the backtests, misses included.
4. Type your own business to run it live. Live runs are rate-limited and have a daily spend cap.

What it does

Step	What happens	Who does it
1. Understand	Turns your words (or a shop link, or a photo of a price list) into a business profile	Nemotron 3 Super, Nemotron 3.5 Super-VL, Tavily Extract/Map
2. Search the web	Finds supply news and local-market news in your country's language, reads the articles, and turns each into a cost or demand shock backed by a quote found word for word on the page	Tavily Search + Extract, Nemotron 3 Super
3. Test plans	An analyst picks 2–6 futures and proposes plans over up to 3 rounds. 12 groups of shoppers react to each plan. A NumPy engine runs 1,000 simulated futures per plan × future, against "Do nothing"	Nemotron 3 Ultra, Nemotron 3.5 Lightning, NumPy engine
4. Your answer	A decision memo: the plan, the bad case (1 run in 20), the cash it ties up, the time to act, the sources, and what isn't modelled	Nemotron 3 Ultra (words), engine (every number)

The rule that makes it trustworthy: the models make decisions (plans, shopper reactions, shock sizes). The engine computes every outcome (demand, revenue, profit, statistics). No model ever does arithmetic.

How we use NVIDIA Nemotron

Every model call in Dry Run goes to an NVIDIA Nemotron model. Each model gets the job its size suits.

Model (Token Factory ID)	Thinking	Job in Dry Run	Why this size	Measured
Nemotron 3 Ultra nvidia/Nemotron-3-Ultra-550b-a55b	on	The analyst. Picks the futures to test (news shock levels, a rival's price move, a new player). Proposes plans, reads the engine's results, and proposes better ones over up to 3 rounds. Can ask Tavily its own question mid-analysis. Writes the memo.	A few calls where deep reasoning is worth the wait	Answer at ≈ 8.5 s, ≈ 340 tok/s. One analysis = 4 Ultra calls ≈ $0.014
Nemotron 3 Super nvidia/nemotron-3-super-120b-a12b	off	The reader. Turns plain text into a business profile, and an article into a shock with a verbatim quote	Strict-schema extraction, where speed matters more than reasoning	1.2–2.0 s and ≈ $0.0003 per profile
Nemotron 3.5 Lightning nvidia/Nemotron-3_5-Lightning	off	The shoppers. 12 kinds of customer each say how they feel about a plan (sentiment mean and spread), shown as pixel faces with a speech bubble	Many small parallel calls	0.94 s TTFT, 198 tok/s
Nemotron 3 Nano nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B	off	Shopper fallback	Kept as backup	5.2 s TTFT, which is why Lightning is primary
Nemotron 3.5 Super-VL nvidia/Nemotron-3.5-Super-VL	off	Reads an uploaded photo or PDF (a price list, a sales report) into the profile	Vision, same API	1.6 s for a 600×400 image

How we call them (see backend/agents/ and backend/services/router.py):

• Out of the box, no fine-tuning. We rely on prompt and system engineering, and the numbers come from the engine, not the model weights.
• Strict json_schema on every call, validated with Pydantic v2 and clamped. Each call retries once, then uses a deterministic fallback. Timeouts are per tier: Ultra 30 s, Super 10 s, Lightning 5 s.
• Thinking is switched per model with chat_template_kwargs. Ultra's reasoning streams in reasoning_content and appears in the UI's analyst notes, while content stays clean JSON.
• A per-model rate limiter reads the x-ratelimit-* headers, honours 429 Retry-After, and drops shopper calls first.
• Every telemetry figure in the UI (Nemotron calls, $ spent) is counted from real call logs.

Shoppers choose with a multinomial logit model that includes a walk-away option. Lightning supplies the sentiment term; the engine supplies the rest:

$$\Pr_{ij} = \frac{e^{\bar U_{ij}}}{e^{\bar U_{i0}} + \sum_{k\in{A,B}} e^{\bar U_{ik}}},\qquad U_{ij} = \beta_{0j} - \alpha_i P_j + \gamma\ln(1+M_j) + \delta,\text{Sent}{ij} + \text{loyalty}{k(i),j}$$

以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗