Case 12 · Graduation Day
BitRouter
More real traction than almost anything a demo day sees — a benchmarked 33% cost cut, three billion monthly tokens, an angel round, named partners — and it buries its best proof (the live Smithers integration) in the Q&A, opens with "Sorry for the trouble," and closes with an ask that has no number.
01 Scorecard · 12 dimensions
02 Klaff 5-lens audit · 2/5
THE ONE REORDER — Delete "sorry," open with the Smithers result as a completed live demo ("While you watched the last talk, BitRouter re-routed its workflow — twenty percent cheaper. That was the demo."), close on a sized ask.
03 Delivery
41 fillers ("uh" ×17, "like" ×15). "Like" clusters almost entirely in the Q&A — the rehearsed pitch is markedly cleaner than the unrehearsed answers.
04 Beat-by-beat
| Time | Beat | What works | The fix |
|---|---|---|---|
| 0:00 | Open | "You guys have Smithers workflow running. Let's try to make it cheaper" — a live callback to the previous talk, the best possible open | Cut "Sorry for the trouble." Then pay the line off immediately with the 20% Smithers result instead of saving it for Q&A |
| 0:10 | Problem | "Token maxing only achieves accuracy by sacrificing cost and latency" — clean thesis, right segmentation | Attach one real bill: what does a production agent loop cost before BitRouter? [confirm a customer number] |
| 0:35 | Solution | "You just set a goal between balancing quality and cost" — the one-knob promise is memorable | The five-item "bring your own…" litany dilutes it. One sentence: works with any model, any workflow |
| 1:05 | Mechanism | RL loop, YAML routing policy, traces → evals → fine-tune-until-goal — a real, describable system | "Set evils to compare against the goal" and "it's our AOS thing" are word salad as spoken [confirm what "AOS" meant]. Two clean sentences, rehearsed |
| 1:35 | Wedge + vision | "We start where token maxing is unsustainable"; "when everybody's loop maxing" is a quotable coinage | Name one live customer or design partner in the wedge — it exists only as a category so far |
| 1:55 | Moat | "Other gateway will rot, but ours will grow with you" — the best line in the pitch; OSS as roots and distribution | Attach the hosted-usage number to the flywheel claim, or the Q&A concession ("run it locally… we don't care") undoes it |
| 2:30 | Benchmark | "Stable thirty-three percent cost reduction without performance loss" on Terminal Bench 2.1 — real and checkable | State the setup (pairing frontier models with Kimi K 2.7) in the pitch, before a technical listener extracts it in Q&A |
| 2:50 | Traction | Three billion monthly tokens · 200 GitHub stars · 2,000+ downloads — real numbers, plural | Triangulate them: whose 3B tokens, hosted or self-hosted telemetry? One sentence kills the doubt [confirm] |
| 3:00 | Team + supply chain | Partnerships named: Tencent, Alibaba, ZAI. "Chinese origin as our supply chain advantage" is a bold, differentiated claim | Name the co-founders; quantify the token discount; prepare for the geopolitical DD thread this line guarantees |
| 3:15 | Ask + close | An ask exists, and the angel round signals momentum | "Potential investor, partner, and contributor" is three asks, none sized. Pick one: raising [$X] to [milestone] |
| 3:25 | Q&A | The two answers hold the pitch's best material: honest eval methodology and the live Smithers integration at 20% | Both belong in the pitch. "You have to give us the trace because we have to get the trace" needs a rehearsed privacy answer |
05 The gaps investors block on
- No business model. Open source + "we don't care" if you self-host + cheaper tokens — three revenue hints, zero revenue sentences.
- The unsized ask. "Investor, partner, and contributor" gives the room nothing to act on.
- The flywheel/self-host contradiction: the moat is traces, but the Q&A says local deployment is fine. Which one is the company?
- Benchmark scope: 33% comes from one configuration (frontier models paired solely with Kimi K 2.7). Learned routing, or one good model substitution? [confirm ablation exists]
- Traction numbers don't triangulate: 3B monthly tokens vs 200 stars and 2,000 downloads — say which measurement. [confirm]
- Team is one spoken name; no co-founders, no backgrounds.
- The China supply-chain claim is unpriced and un-derisked — a genuine edge and a guaranteed DD thread, left as one clause.
Be the routing-intelligence company, not the token broker. The pitch runs two companies: the hosted flywheel whose RL loop compounds on customer traces (OSS as distribution, savings-based pricing), and a token supply-chain business reselling cheaper Chinese-partnership capacity through a gateway storefront. "Other gateway will rot, but ours will grow with you" is the only line a competitor can't say — pick it, price hosted usage and private deployment, and retract "we don't care" about self-hosting.
“BitRouter is the LLM router that gets cheaper every run — drop it into any production agent loop, set one goal, and it learns your workflow until a third of your token bill is gone with zero accuracy loss.”
06 The rewritten pitch
While you were watching the last talk, BitRouter was re-routing its workflow. Same Smithers output — twenty percent cheaper. That's the whole company in one run.
I'm Kelson, co-founder of BitRouter. Everyone here is token maxing — throwing the biggest model at every step — and that's fine until your agent loop hits production, where token maxing buys accuracy by burning cost and latency. BitRouter is the self-improving router: define your workflow's steps in one YAML policy, set one goal between quality and cost, and a reinforcement-learning loop watches your traces, evals every run against your goal, and keeps re-tuning the routing until it gets there. Works with any models, any orchestration, any loop.
Other gateways rot — they route the same on day 400 as day 1. Ours grows with you, because we built observability into the gateway for exactly one purpose: closing that loop. On Terminal Bench 2.1, pairing frontier models with Kimi K 2.7, that's a stable 33% cost reduction with no performance loss — [confirm: ablation vs. plain model substitution] — and we're pushing toward 40–50%.
Today: three billion tokens a month through BitRouter [confirm: hosted vs. telemetry], fully open source, 2,000+ downloads, partnerships with Tencent, Alibaba, and ZAI giving us structurally cheaper capacity. We've raised an angel round; we're now raising [$X] to get [N] production agent companies onto the hosted router and take the benchmark number into their bills. If your agents run in production, give us one workflow — we'll cut it a third.
Brackets are the founder's homework — the copy is done, the proof is theirs to supply. Every claim must match what's actually built.
07 Do these before the next take
- Pick the routing-intelligence identity out loud and write the business-model paragraph — who pays, for what, at what price. This gates the flywheel answer and retracts "we don't care."
- Write the ask — amount, use of funds, one milestone. One sentence of work; currently the room cannot act.
- Promote the Smithers integration from Q&A to cold open — the best live proof in the cohort, told only when asked.
- Run and publish the ablation: learned routing vs. static Kimi substitution on Terminal Bench 2.1, before any technical investor meeting.
- Triangulate the traction numbers in one sentence — whose three billion tokens, measured where. [confirm]
- Name the team — co-founders and one credential each, spoken, every time.
- Rehearse the mechanism section — "set evils" and "AOS thing" are where the word salad lives; the Q&A's "like" ×15 shows what unrehearsed sounds like.