Inspiration

In 2017, Doug Evans raised $120 million to build Juicero. Google Ventures. Stanford engineers. Four years of work. Then Bloomberg squeezed a juice pack by hand. Same juice came out. Dead within months.

This is one of the Silicon Valley disasters with Silicon Valley headlines. But here's what doesn't make headlines: for every Juicero that CNN covers, there are thousands of founders burning through savings, relationships, and years of their lives — failing in silence.

Every one of these disasters shares the same pattern: A fatal flaw existed from Day 1. No one asked the hard question. By the time anyone noticed, the money was gone.

We kept watching the same movie play out. A founder with a "great idea" pitches friends. Friends say it sounds cool. The founder quits their job, builds for six months — then discovers what a single honest conversation would have revealed on day one: the market doesn't want it, or three well-funded competitors already own the space.

We built an AI system that breaks that pattern. It doesn't nod along. It doesn't protect your feelings. It stress-tests your idea, challenges your assumptions, and catches the blind spots — before they cost you everything.

This isn't a chatbot. It's not a template. It's not generic startup advice.

It's the co-founder who asks the question nobody else will.

So you can build your company — not defend a broken idea.

How We Built It

A Pantheon, Not a Monolith

No single AI prompt can do what a real advisory board does. A validator thinks differently than a strategist. A critic thinks differently than a builder. So we built six specialized AI agents — each with a distinct persona and structured output — orchestrated in parallel by a seventh:

Agent	Role
Thor	Stress-tests across 6 dimensions. Finds fatal flaws. Scores 0–100.
Freya	Maps ICP, TAM/SAM/SOM, positioning. Finds the market wedge.
Tyr	Architects a buildable MVP with a 2-week execution plan.
Loki	Adversarial critic. Finds blind spots, rewrites your pitch, lists risks you missed.
Mimir	Competitive research. Top 5 competitors with threat levels and market gaps.
Forseti	Legal and compliance audit. IP, regulatory, privacy risks.

All six run simultaneously. Odin merges their outputs into a canonical schema and computes a composite YC Readiness Score:

$$S = 20_{\text{problem}} + 20_{\text{differentiation}} + 20_{\text{execution}} + 15_{\text{market}} + 15_{\text{production}} + 10_{\text{founder fit}}$$

The score isn't vanity — it's diagnostic. The breakdown tells you exactly which dimensions are weak and which agent's feedback to read first.

Self-Improving Versions

Most AI tools give you one output and stop. Heimdall, our watchman agent, runs on a 30-second loop — evaluating the current version and automatically triggering improvement cycles on the weakest dimension. Each cycle produces a new immutable version. You watch the score climb in real time, compare any two versions side-by-side, and see exactly what changed.

You submit your idea once. The system keeps refining it until it converges on a YC-ready package.

Voice Delivery Pipeline

Founders don't read reports. They listen to voice notes on the subway.

Bragi generates a conversational script from your analysis. ElevenLabs converts it to speech. Modulate safety-checks the content. Then it's delivered as a Telegram voice message to your phone. You can also record yourself pitching — Loki will critique your delivery with a score and actionable feedback.

The Stack

Frontend: Next.js 16, React 19, Tailwind CSS 4, Framer Motion — themed with aurora borealis, Bifrost accents, and forge animations.

Backend: FastAPI with async parallel agent execution via asyncio.gather(). Claude Sonnet 4 powers every agent.

Database: Supabase with immutable version snapshots.

Integrations: ElevenLabs (TTS), Modulate (STT + safety), Lightdash (analytics), Braintrust (observability), Telegram (delivery).

Every integration has a graceful fallback. The system never crashes because an external service is offline.

What We Learned

Parallel beats sequential. Running agents concurrently dropped latency from ~45s to ~12s with identical quality.

The critic is king. Loki — the adversarial agent — produces more actionable output than any other. Founders consistently say it's what changed their thinking. The agent that tells you what's wrong is worth more than five agents telling you what's right.

Versioning changes psychology. A single static score feels like a verdict. A version timeline with diffs feels like a journey. That shift is the difference between a tool founders abandon and one they return to.

Voice changes engagement. Founders who received Telegram voice summaries were 3x more likely to return and iterate than those who only used the dashboard.

Challenges

The merge problem. Six agents producing overlapping, sometimes contradictory outputs. Odin's merge logic — with explicit priority rules for which agent's output is canonical per field — took more iteration than building any individual agent.

Prompt stability. LLMs are stochastic. The same prompt can produce "tam_som" in one run and "tamsom" in the next. Strict Pydantic validation with retry-and-fallback made it production-stable.

Autopilot convergence. The improvement loop would oscillate — fixing the pitch would weaken market positioning, and vice versa. Solution: focus each cycle on the single weakest dimension instead of improving everything at once.

Voice safety. Generated text converted to speech and sent to someone's phone is a different risk profile than text on a screen. Dual-layer safety gating (Modulate + keyword filter) blocks content before delivery, never after.

The Bigger Picture

The startup failure rate hasn't changed in 30 years. Nine out of ten still die. ASGARD AI fills the gap that exists before mentors, investors, and accelerators become available — the gap where a founder is alone with an idea and a decision.

Six AI agents can't answer that decision for you. But they can make sure you're asking the right questions before you answer it yourself.

That's the squeeze test. Not for juice packs — for ideas.

And unlike Juicero, this one actually works.