The struggle is real. So is the data. For interactive website, check out: https://buffet.popped.dev
This repo:
- Reads Telegram group food raid data
- Overanalyzes it with Python, pandas, and unnecessarily pretty graphs
- Tries to predict when and where food will appear, for science (and snacks)
- Renders a gallery of charts about snacks, ice cream, and desperate students
- Useful for data nerds and hungry undergrads
- Python 3.10+ recommended
- Dependencies are listed in
requirements.txt(pandas, numpy, matplotlib, boto3, scikit-learn).
The loader picks a chat export in this order:
BUFFET_CHAT_EXPORT— path to a JSON export file, if setdata/chat_export.json— if presentdata/result.json— if presentdata/chat_export.sample.json— bundled sample- Otherwise it tries S3 via a local metadata API (see
src/load_inspect.py).
For a normal local run, place your export as data/result.json or data/chat_export.json, or set BUFFET_CHAT_EXPORT.
From the project root:
python -m pip install -r requirements.txt
python run_graphs.pyOr use the wrappers (install deps + run the pipeline):
- Windows:
render_graphs.cmd - macOS / Linux:
chmod +x render_graphs.sh && ./render_graphs.sh
On Windows, if you see encoding issues in the console, the scripts set PYTHONIOENCODING=utf-8; run_graphs.py also reconfigures stdout/stderr to UTF-8 when possible.
Outputs are written under out/, grouped by category. A manifest of generated PNGs is saved as out/manifest.json.
K=2 was chosen even though scores keep rising up to K=5 (0.327 → 0.374). The gains beyond K=2 are small enough that the extra clusters would just be hairline splits of the same evening window — not genuinely new patterns. Two clusters is the simplest answer that still draws a real distinction.
The data splits cleanly into a Lunch / early afternoon cluster (38% of posts, peaks 12–14h, strongest on Wed and Fri) and a larger Evening cluster (62% of posts, peaks 19–21h, strongest on Thu and Fri). Evening food is actually the norm — lunch leftovers are the exception.
The lunch cluster is a sharp spike: it rises steeply from 10h, hits ~31% of its posts at 13h, then falls to near-zero by 15h. The evening cluster is a broad plateau — it builds from 16h, sustains ~12% per hour through 17–18h, then crests around 21h. These are two very different behavioural shapes, not just "early vs late".
Both clusters are essentially weekday phenomena. The lunch cluster is spread fairly evenly Mon–Fri with a slight Wed/Fri lean. The evening cluster has a pronounced Friday spike (~32%) and a secondary Thursday bump (~21%) — Friday night events and end-of-week gatherings dominate evening food. Sunday is nearly dead for both.
The side-by-side heatmaps make the two clusters unambiguous. Lunch activity is a tight vertical band from 10–14h running across all weekdays with even coverage. Evening activity is a broader band from 16–22h, densest on Thursday and Friday nights — Friday at 20–21h is the darkest cell in that panel.
Both clusters track the academic calendar: troughs in May–June (exams, then holidays) and December, peaks at semester starts in August and February–March. The evening cluster spikes harder in September (~14%) than the lunch cluster does, suggesting early-semester orientation and welcome events tend to run into the evening. Both clusters are nearly identical in monthly shape, confirming that month is not what separates them — time of day is.
SOL (234) and SOB/LKCSB (223) generate the most food alerts by a clear margin — these two buildings host the highest density of catered events on campus. Connexion (193), SOE (185) and Admin (167) form a second tier. SCIS2 is low (29) because students posting from that building typically write "SOE" instead, since the two schools share the physical building.
Seminar Rooms (SR, 249) appear more than twice as often as Classrooms (CR, 104) — SMU's pedagogy means SRs are the default event space. Lounges (66) rank third, reflecting the volume of casual leave-it-on-the-bench posts. Active Learning Classrooms (ALC, 44) and Function Rooms (44) tie for fourth. Group Study Rooms (GSR, 22) are surprisingly infrequent given how many there are on campus.
Floor 3 is the single biggest hotspot (362 posts), followed by Floor 2 (271). Floors 2–5 account for the bulk of food activity — these are the main teaching floors across most buildings. Basements are more active than expected: B1 alone records 134 posts, mainly from SOL and SCIS seminar rooms. Floor 7 is almost never mentioned (8 posts).
Friday is the dominant food day across every single building. SOL's Friday count (73) is nearly double any other cell in the chart. SOB/LKCSB (63 on Fri) and Connexion (56 on Fri) follow closely. Monday is notably active for SOB/LKCSB (33), suggesting business school events are often scheduled at the start of the week. Sunday is essentially silent everywhere.
Each row is normalised to its own maximum, so small buildings still show their peak shape clearly. SOE has the sharpest lunchtime signal (11–13h), consistent with a culture of midday seminars. SOL and SOB/LKCSB both show heavy evening presence (19–21h). Connexion spreads across lunch and evening roughly equally. LKS Library has a faint signal at 3h — a handful of late-night study-session posts — but its main activity is afternoon and evening.
Food activity across all buildings tracks the SMU academic calendar: troughs in May–July and December, peaks in August–October and February–March. SOL and SOB/LKCSB maintain the largest consistent footprint throughout the dataset. Activity in every building compresses tightly during holiday gaps, confirming the food alerts are driven by on-campus event schedules, not random generosity.
Each building has a characteristic home floor for events. SOA concentrates almost entirely on Floor 2 (61 posts). Connexion is spread across Floors 3, 4 and 5 roughly equally, reflecting its multi-floor event-space design. Admin's hotspot is Floor 6 (60) — the University Lounge — with Floor 5 second (43). SCIS posts cluster on B1 (30), pointing to its basement seminar rooms. LKS Library events predominantly occur on Floor 5 (32), the quiet zone and café level.
Each row is normalised to its own maximum. Seminar Rooms and Classrooms both light up from 10h onwards with a sustained evening peak at 19–21h. Lounges peak twice — at lunch (12–13h) and late at night (21–22h), which matches the "someone left food in the lounge after an event" pattern. Active Learning Classrooms have a strong midday signal and an unusual late-night tail at 23h. The Auditorium shows a distinct 8h spike — a few large morning talks. Training Rooms appear almost exclusively at 7–8h, consistent with early corporate-style sessions.
After you run the pipeline, out/manifest.json lists every PNG path under out/ for that run, and docs/readme/ is refreshed to match.