You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Demo mode with pre-baked messy org. One button on the landing page loads a realistic fake organization: 8 workspaces, a legacy model bleeding money, an obvious caching miss, one underpowered workload. Remark: removes the live-key dependency that kills demos on stage, and lets any judge self-serve the full experience in 5 seconds. This is demo insurance, build it first.
Hero number. Report opens with one giant figure: "€2,340/month wasted," findings ranked underneath as evidence. Remark: judges remember one number and one screen. Everything else is supporting cast. Current layout is dashboard soup, invert it.
Hybrid engine: rules compute, LLM narrates. Keep the 6 rules untouched as deterministic ground truth. Add one Claude call that takes findings JSON plus aggregates and returns structured insights, an executive narrative, and pattern interpretations rules can't encode (workload fingerprinting from input/output ratios and temporal spikes). Remark: the LLM never invents a number, every dollar in the prose is quoted from rule output. This one sentence, said on stage, pre-empts the entire "can you trust it" line of questioning.
Graceful degradation. If the Claude call fails, report renders from rules alone. Remark: your demo cannot depend on an API being up. Also a maturity signal when mentioned in Q&A.
Copy-paste fixes per finding. The exact model string change, the cache_control block, the batch endpoint snippet. Remark: moves the tool from report to action, the line 90% of hackathon projects never cross. "It hands you the patch" is the lean-forward moment.
Tier 2 — credibility layer, ship if Tier 1 is done
Tests on the money math. One test file covering the rules engine and pricing calculations, runnable with npm test, mentioned on stage. Remark: a 17-year-old whose cost math is tested beats a grad team whose demo crashes. This is a Q&A weapon, not engineering hygiene.
Pricing tables stamped and versioned. "Prices as of July 2026" visible in the report footer. Remark: one judge will know current Anthropic pricing by heart. Stale prices silently make every output fiction.
Explainability tags on insights. Each LLM insight carries "based on: input/output ratio, nightly spike pattern." Remark: showing the reasoning trail is the difference between "AI feature" and "engineered system" in judges' eyes.
The ROI footer. "This analysis cost $0.04 and found $2,340/month in waste. ROI: 58,500x." Remark: owns the irony of a token-saving tool spending tokens, gets a laugh, and proves the team thinks in tokens, which is the mission statement in one line.
Deterministic-only toggle. Switch between "rules alone" and "rules + reasoning" views. Remark: a 20-second demo beat that makes the hybrid architecture visible instead of claimed. Also doubles as the fallback UI from item 4, so it's nearly free.
Tier 3 — mission and polish, only after rehearsals are solid
Education expanders. Per finding, a "why this matters" fold-out: what prompt caching is, why batch is 50% off, what makes context bloat expensive. Remark: this is the "give value back, cultivate understanding" goal made concrete, and it scores on impact criteria. Cheap to build, judges love mission framing.
Token literacy score. A 0-100 org-level score derived from the findings mix. Remark: gamification gives the closing line, "we don't just cut bills, we teach teams to think in tokens." Only worth it if it takes under two hours.
Conversational drill-down. Chat grounded in report data: "why is Marketing flagged," "what if we move 60% to batch." Remark: highest wow ceiling, highest time cost, highest crash risk. Build last, cut without mercy if rehearsal time is threatened.
No history/analytics/raw-data investment. Invisible in a 4-minute demo.
No replacing rules with LLM logic. Two runs producing two different dollar figures is instant death. Money math stays deterministic, full stop.
The non-code items, equally mandatory
Rehearse the pitch ten times. Five beats: pain (surprise bill), magic (demo mode, hero number), depth (confidence scoring, honest about aggregate-only data), action (copy the fix), mission (token literacy). Four minutes, timed.
Prepare the three Q&A answers. How do you know a cheaper model suffices without seeing prompts (inference from aggregates, hence confidence scores, limitation turned feature). Why trust the AI layer (it narrates, never computes). What can't the tool see (prompts and content, and we say so in the UI). Remark: hackathons are lost in Q&A more often than in demos.
Whiteboard one rule end to end. Nick should be able to derive the caching-savings math by hand if pushed. Remark: this is the moment that separates "used AI to build it" from "understands what he built."
Write the post-mortem blog post, win or lose. The hybrid-engine design and confidence-scoring approach as a short technical writeup. Remark: at 17, the repo plus the story compounds longer than the prize. This is the actual career asset, don't let it evaporate after the event.
One closing remark on sequencing: Tier 1 is roughly a weekend, Tier 2 another. If the hackathon is close, stop coding at item 10 no matter what and convert all remaining time into rehearsals and Q&A prep. Teams consistently overweight the build and underweight the telling, and this project is already strong enough that the telling is where it wins or loses.
Tier 1 — must ship, these win the hackathon
Demo mode with pre-baked messy org. One button on the landing page loads a realistic fake organization: 8 workspaces, a legacy model bleeding money, an obvious caching miss, one underpowered workload. Remark: removes the live-key dependency that kills demos on stage, and lets any judge self-serve the full experience in 5 seconds. This is demo insurance, build it first.
Hero number. Report opens with one giant figure: "€2,340/month wasted," findings ranked underneath as evidence. Remark: judges remember one number and one screen. Everything else is supporting cast. Current layout is dashboard soup, invert it.
Hybrid engine: rules compute, LLM narrates. Keep the 6 rules untouched as deterministic ground truth. Add one Claude call that takes findings JSON plus aggregates and returns structured insights, an executive narrative, and pattern interpretations rules can't encode (workload fingerprinting from input/output ratios and temporal spikes). Remark: the LLM never invents a number, every dollar in the prose is quoted from rule output. This one sentence, said on stage, pre-empts the entire "can you trust it" line of questioning.
Graceful degradation. If the Claude call fails, report renders from rules alone. Remark: your demo cannot depend on an API being up. Also a maturity signal when mentioned in Q&A.
Copy-paste fixes per finding. The exact model string change, the
cache_controlblock, the batch endpoint snippet. Remark: moves the tool from report to action, the line 90% of hackathon projects never cross. "It hands you the patch" is the lean-forward moment.Tier 2 — credibility layer, ship if Tier 1 is done
Tests on the money math. One test file covering the rules engine and pricing calculations, runnable with
npm test, mentioned on stage. Remark: a 17-year-old whose cost math is tested beats a grad team whose demo crashes. This is a Q&A weapon, not engineering hygiene.Pricing tables stamped and versioned. "Prices as of July 2026" visible in the report footer. Remark: one judge will know current Anthropic pricing by heart. Stale prices silently make every output fiction.
Explainability tags on insights. Each LLM insight carries "based on: input/output ratio, nightly spike pattern." Remark: showing the reasoning trail is the difference between "AI feature" and "engineered system" in judges' eyes.
The ROI footer. "This analysis cost $0.04 and found $2,340/month in waste. ROI: 58,500x." Remark: owns the irony of a token-saving tool spending tokens, gets a laugh, and proves the team thinks in tokens, which is the mission statement in one line.
Deterministic-only toggle. Switch between "rules alone" and "rules + reasoning" views. Remark: a 20-second demo beat that makes the hybrid architecture visible instead of claimed. Also doubles as the fallback UI from item 4, so it's nearly free.
Tier 3 — mission and polish, only after rehearsals are solid
Education expanders. Per finding, a "why this matters" fold-out: what prompt caching is, why batch is 50% off, what makes context bloat expensive. Remark: this is the "give value back, cultivate understanding" goal made concrete, and it scores on impact criteria. Cheap to build, judges love mission framing.
Token literacy score. A 0-100 org-level score derived from the findings mix. Remark: gamification gives the closing line, "we don't just cut bills, we teach teams to think in tokens." Only worth it if it takes under two hours.
Conversational drill-down. Chat grounded in report data: "why is Marketing flagged," "what if we move 60% to batch." Remark: highest wow ceiling, highest time cost, highest crash risk. Build last, cut without mercy if rehearsal time is threatened.
Explicit cuts — do not touch these
The non-code items, equally mandatory
One closing remark on sequencing: Tier 1 is roughly a weekend, Tier 2 another. If the hackathon is close, stop coding at item 10 no matter what and convert all remaining time into rehearsals and Q&A prep. Teams consistently overweight the build and underweight the telling, and this project is already strong enough that the telling is where it wins or loses.