SLM Audit

# SLM Audit (Max 4 Recommendations)

You are auditing this codebase to identify where Small Language Models (SLMs) should be used.

Your goals:

1. Identify existing LLM calls that should be replaced with SLMs.
2. Identify new high-impact AI features enabled by dramatically cheaper, lower-latency SLM inference.

## Step 1 — Scan for LLM Usage

Find all LLM calls (OpenAI, Anthropic, Gemini, Vercel AI SDK, LiteLLM, direct HTTP, internal wrappers).

For each call, briefly note:

* File path
* Model used
* Task type (summarization, classification, extraction, etc.)
* Frequency (approximate)
* Latency sensitivity
* Structured output? (yes/no)
* Logprobs? (yes/no)

## Step 2 — Select Top 4 Total Recommendations

Choose a maximum of **4 total recommendations**, combining:

* Existing LLM calls to replace
* New SLM-enabled product opportunities

Prioritize:

* High frequency
* Latency-sensitive
* Text-in / text-out
* Classification or extraction
* Repetitive patterns
* Structured JSON outputs

## For Each Recommendation Provide:
Feature Name
Type: (Replace Existing | New Opportunity)
Where it runs in product
What it does
Why SLM is appropriate
Estimated volume
Expected impact (cost, latency, product leverage)
Any constraints (logprobs, streaming, etc.)

Rank Recommendations in order of impact.