Logo

Inspiration

Planning an on-campus event means juggling a dozen unknowns: which room fits, how many people actually show up, what it costs, how everyone gets home if the Green Line is down. Each lives in a different spreadsheet or someone's head. We wanted one honest organizer that models each piece from real data and never hides a guess behind a confident number — so an organizer can see not just the answer, but how trustworthy it is.

What it Does

From a single sentence ("300-person overnight AI hackathon at BU, Oct 24–25, budget $15k, needs lunch/dinner/snack/breakfast"), RhettBooked produces:

Venue & rooms — a CP-SAT optimizer assigns the schedule's blocks to real BU event spaces (38 measured spaces) and recommends a building (e.g. GSU).
Turnout — a Monte Carlo over a Beta show-up prior, with the 15-minute arrival/departure curve fitted from real Bluebikes + Numina mobility data.
Food & water — BU dining-hall matching plus a guarded research agent that finds real caterers (with sources) or shows curated real options when offline.
Getting there — nearest MBTA stops (accessibility flagged), parking, and planned disruptions (e.g. a Green Line closure with a fare-free Route 57 shuttle).
People & sponsors — real OpenAlex researchers ranked as potential judges, public professional info only.
Budget — a probabilistic budget (p10/p50/p90) tuned for a student event, with BU-provided resources shown as $0 in-kind.
Three options — Lean / Balanced / Premium, side by side, with live client-side customization sliders.
Drill — a digital-twin stress test (power, occupancy, getting-home) that finds the first failure, proposes fixes, and re-runs to show readiness before → after.

How We Built It

Backend: Python 3.11, FastAPI, pydantic v2, pandas, DuckDB, NumPy, OR-Tools (CP-SAT), scikit-learn, httpx. A provenance layer stamps every output with a label and source and persists it to a provenance table.
Data science (not "an API call"): models produce every number. The arrival curve is a fitted presence profile (running balance of Bluebikes starts−ends, blended 70/30 with Numina pedestrian counts); turnout and budget are Monte Carlo; venue assignment is CP-SAT; people ranking is TF-IDF/cosine over OpenAlex topics.
LLM is optional: backend/app/llm.py is a provider adapter (gemini|none). The LLM only parses the brief and explains results — it never invents numbers.
Guarded research agent + MCP: a read-only MCP server exposes plan_context, osm_places_near, web_search (Gemini + Google Search grounding) and openalex_people. Guardrails are code-enforced: tool allowlist, call/time caps, "no source, no result", OSM cross-check for caterers, and a privacy scrubber for people. Web content is treated as data, never instructions.
Frontend: React 18 + Vite + TypeScript + Tailwind + framer-motion. A VS Code-style docked layout: a full-width main pane with the three plan options and detail cards, and a docked side panel holding the live Customize sliders and a chat-style brief. Light/dark, status shown as colour and text.
Privacy: uploaded registration data is scrubbed (emails/IDs/MAC/phones redacted) and survives only as aggregate counts with cells under five suppressed; HMAC tokenization with a separate vault supports audited re-linking.

Challenges We Ran Into

Honest fallbacks under a rate limit. Our live Gemini key hit its free-tier quota mid-build. Rather than fake caterer data, we built a sanitizer so provider errors never leak into snapshots or secrets, and a curated real-caterer fallback (labelled PLACEHOLDER, real URLs, no invented prices).
A real validation, not a flattering one. We threw out an in-sample "RMSE 0.022" that only measured our own smoothing and replaced it with a temporal holdout (fit on Nov 1–15, test on Nov 16–30).
Realistic cost. The first budget looked like a corporate gala. We retuned the (clearly PLACEHOLDER) rate cards to student-event levels and modelled BU-provided resources as in-kind $0, landing a 300-person hackathon near $5k–$10k.

Accomplishments That We're Proud Of

A temporal-holdout arrival model that beats the naive trapezoid baseline by ~4.6× on RMSE (0.107 vs 0.499) with near-perfect rank correlation (0.974 vs 0.006) on the held-out weekend.
Every number labelled and traceable — a Trust view and /provenance endpoint expose the MEASURED/ASSUMED/PLACEHOLDER/SIMULATED/RESEARCHED mix.
A research agent with code-enforced guardrails, not just prompt text.
Works end-to-end with no API key and no backend (bundled demo snapshots).

What We Learned

Provenance is a feature, not paperwork — labelling forced honest modelling.
"Calling an LLM is not data science": keeping the models in code (CP-SAT, Monte Carlo, KDE/fit) and the LLM to parsing/explaining made the project stronger.
Guardrails belong in code where they can be tested, not in a prompt.

What's Next for RhettBooked

Confirm the PLACEHOLDER BU room/rate data against official BU Reservations.
Blend more months of Bluebikes and BU-specific pedestrian counts into the curve.
Presidio/spaCy NER in the scrubber (names, not just regex patterns).
A draggable panel splitter and per-card deep-dives.

How We Used Kiro

We built RhettBooked with Kiro as the primary development environment, driving the spec-first workflow in .kiro/specs/eventtwin-core and a set of steering rules (.kiro/steering/) for tech stack, product, design system, and hard operating rules (never push without approval, real data only, every number labelled, secrets only in .env). Kiro also runs our research MCP server (see docs/KIRO_USAGE.md): plan_context, osm_places_near, web_search, and openalex_people, so the same guarded tools the app uses are callable from the IDE. Agent hooks run tests on change.

Additional Information

Two submissions, separate repos. RhettBooked (this repo) is the DS+X entry at https://rhettbooked.vercel.app; the frozen Hackachusetts entry is a separate repo.
Reproduce the numbers: docs/NUMBERS.md lists each figure with the exact command. Tests: backend python -m pytest, frontend npm test.
Data sources: BU event spaces (measured), MBTA GTFS, Bluebikes Nov 2024, Boston Numina, OpenAlex, OpenStreetMap Overpass. Room capacities/rates are PLACEHOLDER until confirmed with BU Reservations.