ConferenceMap

Framing the Problem

Every tech conference generates a flood of content — sessions, speakers,

schedules, sponsors — but no AI tool is reliably grounded in it. Ask a

conference chatbot "Who's speaking about React Server Components?" and

you'll often get a plausible-sounding but wrong answer: the wrong speaker,

the wrong room, the wrong time.

This affects everyone at the event. Attendees waste time chasing bad

information between sessions. Speakers have no visibility into what

attendees actually want to know. Organisers have zero signal on which

topics generate confusion or interest. Sponsors rely on random badge scans

instead of qualified conversations.

The root cause: generic LLMs answer from training data or shallow RAG,

with no verification step. Nothing tells you whether the answer is actually

grounded in the event's real data — or whether the model just made it up.

Idea Explanation

ConferenceMap is an AI companion for JSNation & React Summit that answers

questions about sessions, speakers, schedule, and sponsors — and proves

every answer is real.

It's powered by SigMap, a context-verification system we built and

open-sourced (400+ GitHub stars, in production at ING Bank, with published

benchmarks: 78.9% hit@5 retrieval accuracy, 97.9% token reduction).

Instead of dumping the entire conference dataset into the LLM, SigMap

indexes all sessions, speakers, and sponsors, scores them against the

user's question, and assembles only the top relevant chunks — typically

reducing input tokens by ~97%. Gemini 2.5 Flash then generates an answer

constrained to ONLY that verified context, with inline citations.

After generation, we run a faithfulness check: every factual claim

(names, times, rooms) is checked against the source chunks. The result —

input tokens, output tokens, response time, and a grounding score — is

shown live to the user on every single query. This isn't a chatbot. It's

a transparent verification layer.

Built on KendoReact: the Chat interface, session Grid with natural-language

AI filtering, personal agenda Scheduler, and an organiser analytics

dashboard (TileLayout) showing real-time engagement and "coverage gaps" —

questions attendees ask that no session answers.

Implementation

Stack: Next.js 15 (App Router) · KendoReact · SigMap · Gemini 2.5 Flash ·

Google Stitch · Google Cloud Run (europe-west4)

Pipeline:

1. Corpus: JSNation/React Summit sessions, speakers, sponsors, and schedule

scraped into structured JSON (~96 documents, ~15K tokens total)

2. SigMap Index: TF-IDF index built once at server startup, held in memory.

On each query, scores all documents, selects the top 5 above a relevance

threshold, and compacts them to ~300 chars each (~375 tokens total — a

97.5% reduction vs sending the full corpus)

3. Gemini 2.5 Flash: Receives only the verified context plus a strict

system prompt (every factual claim must cite [source: id], or decline

if context is insufficient). Streamed via Server-Sent Events using the

@google/genai SDK

4. Faithfulness Check: After generation, extracts factual claims (names,

times, rooms) from the answer and verifies each against the source

chunks — producing a grounding score shown live in the UI

5. Frontend: KendoReact Chat (premium, 30-day trial) for the main Q&A,

Grid + natural-language AI filter for session browsing, Scheduler for

personal agenda, TileLayout for the organiser dashboard

6. Deploy: Standalone Next.js build in a multi-stage Docker image, deployed

to Cloud Run in europe-west4 (Amsterdam) via gcloud run deploy --source .

UX was prototyped in Google Stitch (4 screens generated from prompts) and

used as the visual reference for the KendoReact implementation.

Challenges

- Token transparency at the right granularity: Gemini 2.5's usageMetadata

includes separate counts for prompt, output, and "thinking" tokens.

Surfacing all three live, mid-stream, required careful SSE event design

so the UI updates progressively rather than only at the end.

- Zero-hallucination by refusal: the hardest part wasn't getting good

answers — it was getting the system to correctly say "I don't have

verified data on that" when no corpus chunk scored above the relevance

threshold, without being overly conservative on borderline queries.

- Real conference data under time pressure: scraping JSNation/React Summit's

actual schedule and speaker data within the first hour was critical — the

whole demo's credibility depends on it being real, not synthetic.

- KendoReact premium components (Chat, Scheduler) require a license —

activating the 30-day trial and wiring the license key into both local

dev and the Cloud Run build pipeline in time.

Accomplishments

- Built a working end-to-end verification pipeline in 24 hours: SigMap

context assembly → Gemini 2.5 Flash streaming → faithfulness check →

live UI metrics

- Achieved ~97% input token reduction vs naive full-corpus prompting, with

a live, animated counter showing the before/after on every query

- Demoed live on JSNation/React Summit's own session and speaker data —

attendees can ask about talks happening in the room they're sitting in

- Shipped 4 full screens (Chat, Session Grid, Agenda, Organiser Dashboard)

using KendoReact's Chat, Scheduler, Grid, and TileLayout components

- Deployed live on Google Cloud Run with a working public URL

Next Steps

- Hybrid retrieval: add embedding-based semantic search alongside TF-IDF

to catch queries where wording differs from the corpus (e.g. "RSC" vs

"React Server Components")

- WebMCP integration: KendoReact's 2026 Q2 release supports registering

components as callable tools for AI agents — letting attendees ask their

own AI assistant to filter sessions and build their agenda directly

- Multi-conference support: generalise the corpus ingestion pipeline so any

conference organiser can plug in their own schedule/speaker data and get

a verified AI companion in minutes

- Persistent analytics: store real query logs (not mocked) to power the

organiser "coverage gap" dashboard with genuine attendee signal over time

- Academic publication: SigMap's benchmark methodology is being prepared

for ICSE 2027 / IEEE Software — this conference deployment becomes a new

real-world evaluation dataset