Explore / Chiku — 30-Day Log Review

What real users did on the ground-transportation prototype, from production logs.
Service next-prototypes  ·  Window: last 30 days (≈ May 9 – Jun 8, 2026)  ·  Snapshot: Jun 8, 2026  ·  Source: Datadog

All numbers below are production traffic unless noted. dev rows (preview/branch & CI pods) are excluded from user-behavior sections. Two surfaces exist: the map-first Explore / Chiku page (where nearly all traffic is) and a separate pservs-backed /search page.

1,033
App events (30d)
940 prod · 93 dev
334
Chiku messages
315 from prod users
$1.28
Total Chiku spend
~$0.0038 / call
169
Trip searches
123 distinct routes
1.9s
Avg Chiku latency
6.1s worst case

What the surface is doing

Traffic is dominated by the Explore / Chiku page. The funnel is visible in the event mix: people set a location, Chiku suggests destinations, places get resolved, and some convert to a trip search.

EventCountWhat it is
llm.prompted334A user message to Chiku
places.resolved233Origin/destination resolved via WAPI
places.predictions193Address autocomplete calls
trips.searched169Actual trip search (MCP)
places.details58Place detail fetches
origin.detected34IP-based "use my location"
search.completed8/search page (dev only)
errors43× WAPI auth + 1× LLM

Chiku (the LLM) — cheap and healthy

claude-haiku-4-5
Model
100% of calls
$1.28
30-day spend
1.06M in / 42.7K out tokens
1,905 ms
Avg latency
max 6,085 ms
1
Hallucinations
1 in 334 (filtered)

Haiku is comfortably handling the task (pick a few destinations + write two sentences). Cost is negligible, latency is reasonable, and the model effectively never returned an off-list destination. Active system prompt: chiku-system-base-v5.

What users typed to Chiku

315 production messages, grouped by primary intent (a message can touch more than one theme).

Intent~MsgsExamples
Setting a location~14084 start with “From …”; bare names (omaha, sf); routes (denver to cheyenne)
Vibe / inspiration~26somewhere warm, Beach weekend, Mountain town, one day trip, foodies
Date / timing~24In August, august 24-27, cheap day trip … today
Vague follow-ups~24More, Anywhere, Yes, What else, Idk
Budget-led~19Cheapest, Round trip under 50, $100 and a free july…
Nearest stop / station~16where will it stop in dc, closest stop to long island
Train / Amtrak mode~14Where can I go by Amtrak from St. Paul, does a massive train go to PA
Off-network / international~10Berlin, Copenhagen, France, Nayarit, northern canada
Logistics / policy~9Traveling with service dog, Does a 6 year old need a ticket, senior fares
Support / UI frustration~7see callouts below

The core loop is working as intended

Location-setting plus open-ended vibe/budget inspiration dominate — exactly what the explore surface is for. “Where am I → somewhere warm / cheap” is the dominant pattern.

Two unmet-need clusters worth a product look

Nearest-stop / station detail (~16 msgs) — users want to know boarding points. Amtrak/rail-specific routing (~14) — one St. Paul user rephrased the same rail question 4 times, suggesting Chiku wasn’t answering it.

Chiku is catching misdirected support & some anti-AI sentiment

“Hello I’m waiting on my bus … haven’t seen no Wanda bus where is my bus”
“Just give me the options like old pen I don’t wanna see the ai”

Lost-passenger, refund, and policy questions land on Chiku with no handoff. A vocal minority wants a plain map/list instead of a chatbot — relevant for the inspiration-page experiment readout.

Where people searched

From the 169 trip searches. Destinations come from a fixed seed list (only 16 ever appear); origins are free-text / geolocated, so they sprawl.

Top destinations

DestinationSearches
New York, NY24
Chicago, IL24
Atlanta, GA15
Miami, FL14
Providence, RI12
Philadelphia, PA11
Washington, DC10
Los Angeles, CA10
Montreal, QC9
Boston, MA9
San Francisco, CA8
Las Vegas, NV8
Cleveland · Detroit · Pittsburgh · Baltimore5 / 4 / 3 / 3

Top origins

OriginSearches
New York, NY (+variants)20
Boston, MA (+variants)18
Albany, NY8
Nashville, TN7
Washington, DC6
Miami, FL6
Syracuse, NY5
Detroit, MI (+variants)7
Tucson4
Omaha, NE4
Hattiesburg, MS4
Denver, CO4
…60 distinct origins total

Most-searched routes

Origin → DestinationSearches
Boston, MA → New York, NY9
Albany, NY → Providence, RI4
Washington, DC → New York, NY3
New York, NY → Providence, RI3
Nashville, TN → Chicago, IL3
Nashville, TN → Atlanta, GA3
Mount Vernon, WA → Chicago, IL3
+ 24 more pairs searched twice, 93 pairs searched once (123 distinct total)

Reliability & the big red flag

168 of 169 trip searches returned zero trips

Across the Explore surface’s MCP search-trips calls, only one (Chicago → Detroit, 5 trips) came back with results. Everything else returned empty. This is the single most important thing to dig into — users are searching real routes and seeing nothing.

/search page: all 8 searches were the same route, all failed

Every search.completed event was BOSSST → NYCPAS (Boston → NYC), returned 0 trips with all 7 carriers failing (GLI, MGL, PPP, BLT, AMT, LML, TDS). These ran only from dev/preview pods, consistent with the VPN/VPC requirement for pservs.

WAPI auth blip — self-resolved

On 2026-06-01 (19:50–19:53) there were 3 wapi.auth_failed timeouts against api.dev.wanderu.dev/v2/auth.json, which caused that day’s 4 autocomplete failures. Nothing since. Plus 1 isolated llm.error (model didn’t call its tool).

Takeaways