Teach an AI to read your datalogs

Why agents are good at this

A 10-minute log at 50 Hz is 30,000 rows across 30+ channels. People scrub through it looking for the obvious spike. An agent that can run code will filter, group and cross-reference all of it in seconds, and it doesn't get bored at row 20,000.

What to log

An agent can only reason about what's in the file. This is the minimum set for fuel and ignition work. Log at 20 Hz or faster for transient work, and at least 10 Hz for steady-state.

Export the log

Every major ECU package can export a log to CSV, and CSV is what agents read best. Native binary log formats usually can't be opened without the vendor's software. The menu labels move around between versions, so the general pattern is:

1. Open the log in your ECU software's log viewer (NSP, PCLink, MTune…).
2. Find Export (usually under File, or a right-click on the log).
3. Choose CSV and include all channels. If there's an option for units in the header, turn it on.
4. Name it with context: 2026-11-02_wot-3rd_E85_18psi.csv.

ECU-specific export notes →

Hand it over

There are two easy ways to give an agent your log:

Then ask a specific question. Here's a good starter prompt:

Attached is a datalog from my [ECU] on a [engine / fuel / boost].
It's a [3rd-gear WOT pull from 2,500 to 7,200 rpm].

1. Load it with code. Tell me the row count, sample rate and channel list.
2. Find the steady full-load section (TPS > 90%) and ignore tip-in and overrun.
3. Compare actual lambda with target. Report every region off by more than 3%,
   with time range, RPM and MAP.
4. Check knock level and ignition angle for timing pulled by knock control.
5. Rank findings by risk to the engine. Show evidence for each.

Don't suggest any changes yet. I want the diagnosis first.

Interactive example

This is our sample WOT pull (a synthetic log, free to download and try yourself). Hover to scrub through it like a log viewer. Toggle the findings to see what a well-briefed agent picks out.

From findings to table changes

Once you agree with the diagnosis, ask the agent to bin the error into your table's cells. Each sample's lambda error is assigned to the nearest RPM × MAP cell. Cells with too few samples are left alone. The result is a map of where the base table is wrong, cell by cell:

I agree with findings ② and ①. For ② only:

- Bin TPS > 90% samples into my VE table axes:
  RPM [4000 to 7000 in 250 steps], MAP kPa [160, 180, 200].
- Shift lambda back by ~0.1 s to allow for wideband lag before binning.
- For cells with ≥ 5 samples, suggest % change = (actual/target − 1) × 100 × 0.5.
- Cap every change at ±4%. Output a table plus a plain-English change log.

Don't touch ignition. Don't apply anything. I'll make the edits myself.

The script approach

Once you've found a workflow you like, ask the agent to save it as a script. Every future log then gets the same consistent analysis, and you can see exactly what it did. Here's a cut-down version of our free log_summary.py:

import pandas as pd

df = pd.read_csv("sample-wot-pull.csv")
wot = df[df["Throttle Position (%)"] > 90].copy()
wot = wot.iloc[8:]                       # skip the tip-in transient
wot["err_%"] = (wot["Lambda 1"] / wot["Lambda Target"] - 1) * 100

# flag sustained lean: > 3% off for 5+ consecutive samples
lean = wot["err_%"] > 3
runs = (lean != lean.shift()).cumsum()
for _, g in wot[lean].groupby(runs[lean]):
    if len(g) >= 5:
        print(f"LEAN {g['Engine Speed (rpm)'].min():.0f}-{g['Engine Speed (rpm)'].max():.0f} rpm "
              f"peak {g['err_%'].max():+.1f}%")

knock = wot[wot["Knock Level (V)"] > 0.8]
if len(knock):
    print(f"KNOCK at {knock['Engine Speed (rpm)'].min():.0f} rpm, peak {knock['Knock Level (V)'].max():.2f} V")

Pitfalls to teach your agent

These are the mistakes agents (and new tuners) make most often. Put them in your engine brief so the agent checks for them every time.