From Data to Evals

A practical guide to from data to evals for AI founders.

What This Lesson Covers

From Data to Evals is a key topic in AI Telemetry & Observability. In this lesson you will learn the underlying principle, why it matters specifically for AI startups, the playbook experienced founders use, and the patterns to avoid. By the end you will be able to apply from data to evals on your own startup with confidence.

This lesson belongs to the Product & Engineering category of the AI Startup track. AI startups succeed or fail on the same things every startup does — clarity of customer, defensible moat, focused execution — plus AI-specific dynamics around model dependency, talent wars, and rapid platform shifts.

Why It Matters

Instrument AI features so you can debug and improve them. Learn LangSmith/Helicone/Langfuse patterns, the metrics that matter, and the data-collection consent loop.

The reason from data to evals deserves dedicated attention is that the difference between an AI startup that becomes a category leader and one that gets stuck at $1M ARR usually comes down to a small number of decisions made early. Two teams with the same idea can end up in very different places based on how well they execute on this. The patterns below are taken from the founders who got there first — learning them does not guarantee the win, but skipping them almost guarantees a slower path.

How It Works in Practice

Below is a worked example of how to apply from data to evals in a real AI startup context. Read it once, then sketch out how you would apply it to your own situation.

# AI telemetry pattern with LangSmith / Langfuse / Helicone
from langsmith import traceable
from openai import OpenAI

client = OpenAI()

@traceable(name="answer_question", run_type="chain", tags=["prod"])
def answer(question: str, user_id: str) -> str:
    docs = retrieve(question)            # @traceable too
    response = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[
            {"role": "system", "content": f"Answer using context:\n{docs}"},
            {"role": "user", "content": question},
        ],
        metadata={"user_id": user_id, "feature": "qa"},
    )
    return response.choices[0].message.content

# Capture in-product feedback (thumbs up/down) and link to trace ID
# Top-down: review the worst 1% of traces every week
# Bottom-up: add the worst examples back into your eval set

Step-by-Step Walkthrough

1. Anchor on a real-world example — Pick one AI startup whose execution of from data to evals you admire. Study what they did and the trade-offs they accepted.
2. Define your inputs — Get the data, customers, dollars, or commitments you need before deciding. Decisions made without inputs are guesses.
3. Pick the smallest reversible step — Most decisions can be tested before being committed. Find the cheapest test that produces real signal.
4. Set a kill criterion in advance — Decide what would tell you to stop, BEFORE you start. Without it, sunk-cost fallacy will keep you in.
5. Communicate the decision and reasoning — Write it down. Future-you and future hires will need to know what you decided and why — not just what you did.

When To Use It (and When Not To)

From Data to Evals is the right move when:

• The decision is non-trivial AND the consequences will compound
• You have enough data (customer signal, financial information, team feedback) to decide responsibly
• You can commit the team and capital required to execute
• The risk of inaction is greater than the risk of moving forward

It is the wrong move when:

• A simpler, cheaper decision would meet the need
• You do not yet have the inputs needed to decide responsibly
• The decision can be deferred until you have more signal
• You are still iterating on the underlying strategy — commit to the strategy first

Founder Checklist

• Have you reduced the decision to one sentence you could explain to a non-founder?
• Do you know the cost of being wrong (in dollars, time, talent, market position)?
• Have you discussed the decision with a peer founder, an advisor, OR a coach?
• Have you written down the decision and the reasoning so you can revisit it in 90 days?
• Have you set a kill criterion you can recognize without ego getting in the way?
• Are the team members affected aware of the decision and the why?

Next Steps

The other lessons in AI Telemetry & Observability build directly on this one. Once you are comfortable with from data to evals, the natural next step is to apply the patterns from the surrounding lessons — that is where compound returns kick in. Startup decisions are most useful as a system, not as isolated tactics.