Hack Day Starter
Architecture & AlgorithmDeterministic Engine

How Hack Day Starter Works

Why we use reproducible mathematical algorithms rather than probabilistic LLMs to size hardware, and how our recommendation weights are computed.

1. Deterministic Rules Beat LLM Hallucinations

When choosing an AI model for a weekend hackathon or production laptop, an AI chat prompt like "What model fits on my 16GB Mac?" often hallucinates nonexistent model variants, invents inaccurate parameter counts, or recommends 32B models that trigger instant kernel panics.

Hack Day Starter uses a 100% deterministic, test-backed engine. Given the same hardware inputs (RAM, GPU type, VRAM, and OS), the engine computes the exact same reproducible ranking every single time.

2. The Two-Stage Pipeline: Safety Gates & Scoring

1Stage 1: Hard Safety Gates (Binary Elimination)

Before any score is calculated, every model candidate must pass five non-negotiable safety criteria:

1. Verification Status

Must be verified against official primary sources. Discrepancies and stale models are blocked.

2. Active Lifecycle

Must not be retired or deprecated from practical use.

3. Free Disk Space Buffer

Must satisfy: Free Disk ≥ Model Size + 1.5 GB to allow temporary blob unpacking.

4. Physical Memory Ceiling

Model weight plus runtime buffer cannot exceed physical RAM or available GPU VRAM.

2Stage 2: Weighted Compatibility Scoring (Max 100 Pts)

Models passing all hard gates are evaluated across four weighted dimensions:

Use Case Alignment (Up to 40 Points)

Rewards models whose verified strengths match the user’s declared workload (code, chat, summarization, general reasoning).

40%
Memory Headroom Optimization (Up to 30 Points)

Calculates the sweet spot between running a model large enough to be intelligent, but leaving sufficient headroom to prevent system memory paging.

30%
Hardware Acceleration (Up to 20 Points)

Awards bonuses when models can be 100% offloaded to NVIDIA CUDA Tensor Cores or Apple Silicon Metal unified memory.

20%
Tool-Calling Capability Bonus (10 Points)

Awards additional weight when models possess verified native function calling, enabling agentic workflows.

10%

3. The 1.5 GB Safety Buffer Rationale

When Ollama pulls a model via ollama pull <tag>, it streams compressed tarballs from the registry and extracts the layer blobs to ~/.ollama/models/blobs. If the target drive has zero remaining bytes during extraction, the download crashes and leaves orphaned temporary files.

Similarly, operating systems (macOS WindowServer, Windows DWM, Linux systemd) require free RAM to handle display composition and background networking. Our mandatory 1.5 GB buffer policy ensures neither disk writes nor active RAM allocations risk crashing the host environment.

← Back to Hardware ProfilerView Model Registry →