B

Beakn

A Pinterest-style hub that holds the content, a studio that creates it, and an AI that answers any question.

Designed by Sandeep Karnati

0
smart-home products
0+
app features
0
kinds of content created

scroll · or use Next ↓

The problem

Making content is slow.
Finding the answer is slower.

The usual way

  • 🎨 Every image, page & video made by hand
  • 📄 Specs scattered across PDFs & sheets
  • 📞 Every "does it support…" goes to support
  • 🤷 No record of what customers keep asking

Beakn

  • 🪄 A studio generates 5 kinds of content from a reference
  • 🧱 One hub holds it all, Pinterest-style
  • 💬 An AI answers instantly, with clickable sources
  • 📈 Every question teaches the catalog what's missing

How it all fits

One loop, end to end — tap each step

📦
Catalog
products & features
📚
Knowledge
specs → knowledge bank
🎨
Create
studio: 5 media
💬
Answer
web + WhatsApp
📈
Improve
gaps → content

👆 Tap a step to see what happens there.

Feel it

A customer asks in plain words → a grounded, cited answer

💬 a customer, typing:

✨ the answer, grounded in the knowledge bank:

Yes — 16A relay per channel handles 1.5-ton ACs (7–10A draw) [1]

🔌 No neutral required — works with both wirings [1]

🛡️ 4KV surge protection, isolated SMPS [2]

— Source: LPF 6 Switch · Specs · FAQ (tap to open)

Chapter 2

A Pinterest-style hub

A masonry feed of every product and feature — videos, photos, reels, files and live pages, flowing in gap-free. Browse freely; sign in only to ask.

Every room of the hub — tap one

👆 Tap a room.

It all feeds one knowledge base — which the studio writes and the AI answers from.

Chapter 3

Five kinds of content, one studio

Images, web pages, articles, videos and reels — each generated for your real product, with its own set of creative controls. Tap a type.

For the curious — what "decode" means
Each medium has an editable decoder (a master prompt). You hand it a reference you admire — an image, a page, an article, a video or a reel — and a vision/text model reverse-engineers ONLY its craft: an image's scene, lighting & product staging; a page's real HTML/CSS design tokens; an article's tone & structure; a video or reel's shot flow & hook. The reference's actual subject and words are never reused. The output is a reusable prompt with {{product}} / {{feature}} slots, filled later from your product's real photo and knowledge bank. Images can be made with Imagen or GPT-Image; videos render with Veo.

Chapter 4

How the studio turns inspiration into your product

Decode a reference once, then apply it to any product, forever. Tap each step.

🔮
Decode
a reference you admire
🎛️
Capture
a reusable master prompt
🎯
Apply
to a real product
Generate
→ saved to media

👆 Tap a step in the studio pipeline.

🎯

Your real product

It never invents the device. A vision model writes a "visual spec" from the actual product photo, then that photo is handed to the image model as a reference — so it renders YOUR product, not the reference's.

🎚️

Deep creative controls

Per-medium fields — occasion, style, shot, lighting, hook, format, layout… — each backed by a researched phrase library so a tap becomes a rich instruction.

🔑

SEO built in

Articles & pages pull live keywords (DataForSEO, India/English) scoped to the product & category, with a one-click remove for competitor terms.

🏷️

Brand control

Place the Beakn logo cleanly in a corner (composited for real), or generate logo-free — with optional designed captions when you want text.

🎲

Variation strength

Off → subtle → medium → wild: sample fresh style modifiers so the same recipe yields a different, on-brand take every time. One-tap reroll.

🎬

Real video generation

Videos aren't just scripts — they're rendered for real with Veo, polling until the finished clip downloads to the media library.

For the curious — admin owns the engine
Every decoder is versioned, seeded on first use, and restorable to default. Admins edit the master prompts, the per-medium field vocabularies and phrase library, the variation pools, and the model routing (which model decodes, generates and writes each medium) — all from a Studio config console, no code. A separate webpage generator returns a full self-contained responsive HTML document you can open in a new tab.

Chapter 5

Save it once as a recipe

A great prompt shouldn't be a one-off. Capture it, name it, star it, and re-cook it for any product on demand.

📒

Recipes

Any decoded or hand-tuned prompt is saved as a recipe with its medium, facet tags, reference thumbnail and a use-count — then re-applied to a different product in one tap, filling that product's real context.

📚

Library & compare

A searchable, paginated library of recipes and creations — filter by medium, product, starred or facet; sort by newest, most-used or A–Z; and compare prompts side by side.

🔣

{{ Placeholders }}

Prompts use {{product}} / {{feature}} / {{category}} variables; missing values are cleanly stripped so a raw token never reaches the model. Open straight into ChatGPT, Gemini, Claude, Midjourney, DALL·E or Sora.

🗃️

Creations history

Every generation is kept with its product, recipe and settings — approve or keep as draft, delete, or turn an image straight into a card on the product. Output is saved to media with a reference link.

🛠️ There's also a standalone Prompt Library for hand-written prompts — categorised (image, video, social, copy, ad), starred, with tool deep-links that pre-fill the prompt where supported.

Chapter 6

From specs to a searchable brain

Before the AI can answer, every product's knowledge is distilled and indexed. This is how.

Knowledge banks

📝

Meta summary

A tight 200–300 word overview of each product, category and feature — the fast facts.

📖

Complete write-up

A full document (overview, specs, features, use-cases, installation, FAQ, compatibility…) generated from the category source + specs, with versions.

👁️

Visual spec

A vision model describes the product from its own photo, so generated content and answers match the real device.

Drafts can be seeded from a product's cards — or extracted from uploaded PDF & DOCX spec sheets — then edited and published.

The indexing pipeline — tap each step

🗂️
Gather
KBs + cards
✂️
Chunk
by heading
🧠
Embed
meaning vectors
🗄️
Store
tagged in Qdrant

👆 Tap a step of the pipeline.

The knowledge console

📊 Index status

See what's indexed per product, category and feature — never / stale / indexed, with last-indexed dates — and re-index everything or just one product in a click.

🧩 Chunks browser

A dedicated tab to browse every indexed chunk, with search, filter by source, sort and pagination — so you can see the AI's raw knowledge.

⭐ Human-corrected answers

Edit any answer in the Conversations log; it's pinned as a canonical source that survives every re-index — a quiet "human corrects the AI" loop.

🕳️ Content gaps

Questions the knowledge base can't answer are logged as gaps in Analytics — a to-do list of content to create next.

For the curious — the mechanics
Sources (product / category / feature knowledge banks + content cards) are split into ~375-token chunks along their headings, embedded in batches with OpenAI text-embedding-3-large (3072-dim), and upserted into a self-hosted Qdrant collection (and mirrored to a chunks table) with payload tags (productId, categoryId, featureId, sourceType, sourceRef, title, section). Re-index runs for the whole catalog or a single product; a status query reports counts and freshness. The serving cache is kept separate from durable conversation history, so re-indexing never deletes a customer's saved questions.

Chapter 7

How a question becomes a trusted answer

A self-hosted RAG pipeline — running on your own server, not a black box.

The pipeline — tap any block

💬
Web + WhatsApp
a floating Ask button scoped to the product · or WhatsApp · same brain
🧠
Embed the question
a 3072-dim meaning-fingerprint
🔍
Hybrid search
meaning + exact keywords
🎯
Rerank
a self-hosted cross-encoder
Instant cache
repeats served with no model call
✍️
Grounded, cited answer
streams token-by-token · cites sources · refuses to guess

👆 Tap any block of the pipeline.

Watch retrieval happen — tap a question

A simulation of how the right knowledge lights up.

📦 LPF 6 Switch · Specs
✨ Scenes feature
🗂 LPF Retrofit · Category
📦 LPF 6 · Installation
❓ LPF 6 · FAQ
📦 Modular Touch 4 · Specs
✨ Schedules feature
🗂 Energy Meters · Category
For the curious — the retrieval stack
A question (condensed with the recent conversation for follow-ups) is embedded, then runs hybrid retrieval — a dense vector search (top 25) and a keyword search (top 15) in parallel — so exact specs like "16A relay" are never missed. A self-hosted bge-reranker-v2-m3 cross-encoder re-scores the candidates and keeps the top ~6, which become the answer's context. Per-product chats filter to that product + its category + always-global features; the global chat searches everything and recommends the best product. If nothing scores strongly enough, the assistant declines instead of guessing.

Chapter 8

Answers you can trust — and check

Every claim comes with its source, and a human is always one tap away. Watch an exchange replay:

📦 LPF 6 Switch 🌐 Whole catalog
Does the low-profile switch need a neutral wire?
🔎 searching the LPF 6 Switch knowledge bank…
📦 LPF 6 · Specs ❓ LPF 6 · FAQ
Talk to a human →

🧾 Clickable sources

Only the chunks the answer actually used are shown — numbered [1][2], each linking straight to that product page.

👍 Feedback & escalation

Thumbs-down answers and "talk to a human" requests land in a support inbox with the full transcript.

🚫 Never invents

If the knowledge base has no answer, it says so and logs a content gap — instead of guessing.

⚡ Instant repeats

A semantic cache recognises the same question asked differently and answers in a blink — no model call.

🔁 Multi-turn

Follow-ups like "and the 4-channel one?" understand the conversation so far.

🔐 Fair access

An account is required to ask; customers get a daily question allowance; staff are unlimited.

Chapter 9

The right brain for the right job

Some brains run on your own server; the writers are swappable from Settings — and you pick the best by comparing them.

📚

The Librarian — embeddings + Qdrant

Turns every knowledge chunk and every question into a meaning-fingerprint, stored in a self-hosted vector database.

🎯

The Judge — a self-hosted reranker

A cross-encoder that re-reads the candidate chunks against the question and keeps only the truly relevant ones. Runs on your hardware.

✍️

The Writers — Claude · GPT · Gemini

The models that write answers and content — any of them, routed per purpose (chat, knowledge, studio) through one gateway, switchable without touching code.

👁️

The Eyes & Makers — vision · Imagen · Veo

A vision model describes each product; image models render it; Veo renders the videos. Provider keys are stored encrypted.

⚖️ Compare mode: run one question through several writers side-by-side, see each answer's speed and cost, and lock the best as the default — with per-request token & cost tracking (live OpenRouter pricing) and today / month spend dashboards.

For the curious — the actual stack
Embeddings: OpenAI text-embedding-3-large → self-hosted Qdrant. Reranker: bge-reranker-v2-m3 served as a small PyTorch/CPU FastAPI microservice (the standard inference image crash-loops on this AMD host). Answer models: Claude, GPT, Gemini and open models routed through OpenRouter (plus direct OpenAI/Anthropic/Google), assignable per purpose with encrypted keys. Image: Imagen / GPT-Image; video: Veo (async). The answer cache lives in pgvector beside the app database. Every model call's tokens and real cost are logged for the spend dashboards.

Chapter 10

For customers and for the team

The same brain, surfaced everywhere it's needed. Tap a surface.

For customers

For the team

👆 Tap any surface above.

Chapter 11

Built by AI — held to engineering discipline

Every feature shipped through an enforced pipeline. Nothing is "done" until it's verified live.

🎫
Linear ticket
one flat issue per task
⌨️
Code
one branch, focused
🧪
Build + smoke
red = nothing ships
🚀
Deploy
all-or-nothing gate
Verified live
then — and only then — done

Tracked in Linear · versioned on private GitHub · deployed behind a Cloudflare tunnel · images auto-optimised to tiny WebP (a 7 MB render → ~15 KB).

From a catalog to a brain — chapter by chapter

The hub

A Pinterest-style feed of every product & feature — cards, profiles, categories, boards, reels, packs.

It learned to create

The Creative Studio: decoders for images, pages, articles, videos & reels, with recipes, SEO & real generation.

It gained knowledge

Per-product knowledge banks (meta + write-up + visual spec), then indexed into a vector store.

It grew a memory

A self-hosted RAG stack — Qdrant + a cross-encoder reranker + embeddings + hybrid search.

It started answering

Grounded streaming answers with citations, a semantic cache, compare mode, and cost dashboards.

It reached out

Semantic search, an AI product-compare, feedback & escalation, My Questions, and a WhatsApp assistant.

Most catalogs list what a product is.
This one creates its content and answers what a customer needs.

Content, creation and support — one hub, one brain, getting smarter with every question asked.

Open Beakn →