Text here to Interact with all the Agentic Media Pipelines
Agentic Pipelines orchestrated with Genblaze
Killer Product Promos generated with Seedance+Gemini-2.5-flash
GIF
Automated Product Demos orchestrated by Gemini-2.5-flash+Firecrawl+LMNT
Brandkits generated with Gemini-2.5-flash+Gemini-2.5-flash-image
Appkits for Desktop and Mobile generated by Gemini-2.5-flash+Gemini-2.5-flash-image
Social Listening (Reddit Reachouts) performed by Gemini-2.5-flash+Gemini-2.5-flash-image+Tavily+Jina
VCs/Investors Outreach powered by Gemini-2.5-flash+Gemini-2.5-flash-image+Exa Search
Competitor Analysis done withGemini-2.5-flash+Gemini-2.5-flash-image+Jina
Pitch Deck Generations with Gemini-2.5-flash+Gemini-2.5-flash-image
All generations stored and served from Backblaze

Inspiration

We kept watching the same scene play out: a solo founder ships a real product at 2 a.m., then spends the next two weeks drowning in Canva tabs, competitor spreadsheets, half-finished demo videos, and the classic “I’ll design the brand later.” Building got cheap. Looking like a company did not.

The one-person billion-dollar company stopped sounding like a meme and started sounding like a business model, except the non-engineering half of the company still demanded agencies, retainers, and six-week cycles. We wanted the opposite of another AI writing assistant that spits LinkedIn posts. We wanted files you can ship: a cinematic promo video, a brand-kit ZIP, a UI kit pack, an automated demo with narration, a social-listening PDF, a competitor analysis PDF, an investor brief, and a short on-brand pitch deck — the operational output of a team you cannot hire yet.

That asymmetry is what lit the fuse. If code generation collapsed the cost of engineering, something should collapse the cost of brand, product surface, GTM, and fundraising ops. FounderBlaze is our answer to that gap.

What it does

FounderBlaze is an entire collection of departments in a typical startup, distilled into eight AI services, routed by one entrypoint agent:

Promo Video — product URL → cinematic ad (Segmind Seedance)
Automated Product Demos — live site + script → narrated screen automated product demos for your social media and clients
Social Listening — landing page → Reddit threads where people and orgs already need what you sell, plus draft replies so you can jump into the conversation (PDF)
Outreach — website + revenue sheet → investor intelligence PDF with their contact information
Competitor Research — product name → competitive analysis PDF
Brand Kit — name + brief → logos, palette, fonts, banners (ZIP)
App Kit — name + idea (+ optional brandkit file) → desktop/mobile UI kits mock (ZIP)
Pitch Deck — product URL + funding ask → researched, design-matched 6–8 page investor deck (PDF)

Founders chat with a Gemini agent (or call MCP / HTTP). Required fields are gated so the agent cannot invent URLs. Jobs run asynchronously via A2MCP → Temporal → Genblaze pipelines. Deliverables land on Backblaze B2 as downloadable artifacts — not dashboard fluff.

In spirit: one request in, one specialist multi-agent pipeline out, one file you can open and view.

How we built it

We treated FounderBlaze as infrastructure, not a demo script farm.

Runtime brain — Genblaze:

Every service is a Genblaze Pipeline of SyncProvider steps (TEXT / IMAGE / VIDEO) with fan-in via input_from. Gemini text/image goes through genblaze_google; B2 upload goes through Genblaze’s ObjectStorageSink backed by genblaze_s3. Temporal activities call run_*_pipeline and map Genblaze step events into heartbeats — the worker schedules work; Genblaze is the work.

Durable last mile — Backblaze B2:

Objects live under founderblaze/{service}/…. After each run we finalize Genblaze provenance (pointer for video, sidecar for PDF/ZIP) and merge it onto job artifacts. The API re-signs download URLs on poll. No alternate primary store.

Surface area. FastAPI A2MCP gateway (:4021), Temporal worker, Gemini entrypoint agent + MCP (:4022), Next.js chat UI. One catalog of tools powers chat, agent HTTP, and MCP — same validation, same jobs.

The eight Agentic Pipelines:

1. Promo Video
Product URL goes in → Firecrawl maps the site and grabs screenshots → Gemini writes the ad concept + Seedance prompt → Seedance 2.0 generates the MP4 (audio baked in) → final video uploads to B2.

2. Automated Product Demo
Live site URL + natural-language script go in → Gemini plans atomic browser steps → Firecrawl + Playwright record the real product → LMNT speaks the narration → ffmpeg muxes narrated MP4 → uploads to B2.

3. Social Listening
Product URL / context goes in → Gemini understands the product → Tavily finds relevant Reddit threads → Gemini drafts compliant replies → charts rendered into the report → PDF compiles and lands on B2.

4. Outreach
Website + revenue spreadsheet go in → sheet downloads and parses → Gemini + Exa analyze the company and revenue → Exa finds investors, portfolio comps, and partner contacts → contacts enrich → charts generated and embedded → intelligence PDF uploads to B2.

5. Competitor Research
Product URL goes in → search (Serper → Brave → DuckDuckGo) finds competitors → evidence scrapes via Jina/Gemini → feature diff, pricing scrape, positioning/SWOT → charts embed → analysis PDF uploads to B2.

6. Brand Kit
Brand name + brief go in → Gemini invents concepts → Gemini image generates logo options → palette + fonts lock in → visuals, icons, and banners generate → everything zips (manifest.json included) → ZIP uploads to B2.

7. App Kit
Product name + idea go in (+ optional brand-kit ZIP) → Gemini plans 4–6 screens → brand context resolves (from ZIP or invented) → Gemini draws two 16:9 UI kit boards (phone + desktop, all screens on each) → boards zip → ZIP uploads to B2.

8. Pitch Deck
Product URL + funding ask go in → Gemini researches the product (search-grounded) → design language extracts from the live site → Gemini plans exactly 6–8 slides → Gemini generates landscape 16:9 slide images → Pillow compiles the multipage PDF → PDF uploads to B2.

Shared surface

FastAPI A2MCP (:4021) + Temporal worker + Gemini agent/MCP (:4022) + Next.js chat. One tool catalog, same validation, same jobs. Gemini text/image via FounderBlaze chat_with_retry / generate_content_with_retry; agent on gemini-3.1-pro-preview.

If you squint at cost asymmetry the way we did, it looks something like:

$$ C_{\text{agency}} \gg C_{\text{pipeline}} \quad\text{while}\quad T_{\text{agency}} \gg T_{\text{pipeline}} $$

We built for the right-hand side: minutes, not weeks; files, not retainers.

For more information about the Technical Implementation, visit our EXTREMELY detailed technical readme over here: https://github.com/Marshal-AM/founderblaze/blob/main/README.md

Challenges we ran into

Making eight services feel like one product. Each pipeline has different modalities, SLAs, and failure modes. Forcing them through one job envelope, one tool schema, and one provenance story took more discipline than wiring any single demo — including Pitch Deck’s hard 6–8 page contract and image→PDF compile path.

Genblaze + B2 as a contract, not a bolt-on. Seedance and other vendors return transient URLs; founders need a durable object. Emitting only the final asset through ObjectStorageSink, then attaching provenance sidecars without corrupting hashes, was fiddly — especially across video (pointer) vs PDF/ZIP (sidecar) modes.

Rate limits and long fan-outs. Image-heavy paths (brand logos, app-kit desktop+mobile per screen, pitch-deck slides, insight charts) force pauses, retries, and explicit aspect-ratio configs so Gemini does not default to square “slides.”

Accomplishments that we're proud of

All eight services actually ship. Promo Video, Automated Product Demo, Social Listening, Outreach, Competitor Research, Brand Kit, App Kit, and Pitch Deck each run end-to-end and hand back real deliverables — cinema-ready ads, narrated demos, research PDFs, brand ZIPs, UI mock packs, and 6–8 page investor decks — not stubbed JSON or “coming soon” placeholders.
The whole system is agent-friendly. We exposed every pipeline through MCP so founders (and other agents) can plug FounderBlaze into any agentic stack — Cursor, Claude Code, custom MCP clients — without rebuilding our orchestration. Same tools, same validation, same B2 artifacts, whatever agent is driving.
Genblaze + Backblaze as load-bearing, not marketing stickers: every completed job ends in a B2 artifact with provenance.
Required-field gating that treats invented inputs as a bug, not a feature.
A monorepo that can run locally end-to-end: gateway, worker, agent, chat — and CLI live runners for each pipeline.

What we learned

Time is the scarce resource for early-stage founders. Generic agent setups can eventually produce brand kits, research memos, pitch decks, or demo scripts — but they often take hours of back-and-forth prompting, tool babysitting, and stitching. Founders cannot wait that long when a launch window, investor reply, or cold outreach thread is open now.

That is why FounderBlaze pipelines are built to return a finished artifact in minutes — most services around five minutes; image-heavy decks like Pitch Deck still finish as a downloadable PDF instead of a week of deck design:

$$ T_{\text{pipeline}} \ll T_{\text{agency}} $$

Specialist Genblaze graphs beat open-ended agent loops when the job is clear: get the file, move on.

Agent-native products need hard contracts. Soft prompts are not enough when money and hours sit behind a tool call. Schemas + message grounding beat “please don’t hallucinate.”

Pipelines beat prompts. Specialists with modalities, fan-in, and sinks beat one mega-prompt that tries to be a designer, researcher, and fundraiser at once.

Vendor glue is the job. Gemini, Seedance, LMNT, Exa, Tavily, Firecrawl, Jina, search failover — the art is orchestration and failure modes, not any single model card.

What's next for FounderBlaze

Instagram / TikTok-style AI influencer content.
Input: your product URL.
Output: FounderBlaze spins up a persistent AI creator who becomes the face of your growth. Every day, a fresh multi-agent pipeline researches trends, mines your product for new angles, writes a brand-new script (never repeated, never stale), and renders a ready-to-post Reels/TikTok video — hook, voice, face, captions, all done.
Why next: distribution is the bottleneck after you have a brand and a product surface. Founders need daily short-form content without hiring a creator team or burning evenings in CapCut.

Documentary-style founder story video.
Input: the founder’s headshot image.
Output: a 2–3 minute cinematic origin-story video — narration, B-roll-style AI visuals, music, captions — the kind used for crowdfunding or investor emotional hooks.
Why next: fundraising and crowdfunding still hinge on why you as much as what you built. A polished founder story is usually an agency job; we want that file in the same minutes-scale delivery model as the rest of FounderBlaze.