Tadah. AI-powered content accessibility.

The thesis

Tadah started with a single observation. The web has become unreadable for most of the people on it. Not because content is hidden, but because it does not meet readers where they are. A health article reaches you in English at a grade-twelve reading level, and you need to share it with parents who do not read English fluently. A YouTube video about farming technology runs in a language a Nigerian farmer does not speak. A tweet thread on responsible AI sits behind enough jargon to lose the audience it was meant for. Translation tools exist. Summarizers exist. Text-to-speech exists. No product had stitched them together with the recipient in mind.

We set out to build content alchemy. One platform that could take any source, in any format, and produce a derivative tuned to the consumer: their language, their complexity tolerance, the format they would actually consume on the device they actually held. Audio for screen-tired commuters. Pidgin English for diaspora parents. Eighth-grade Spanish for the cousin who never finished school. The product brief called it “personalized derivatives from any source content, to make it more accessible for underserved users and those in their social graph.”

What we shipped

The first cut of Tadah was deliberately narrow. Users could send any article URL through three different channels, and Tadah returned a short audio summary with background music, a transcript, and a link back to the source. The three channels mattered more than they look on paper.

1. Email. Forward or send any URL to a Tadah address, get audio back. Mailgun handled inbound. A regex parser pulled the URL. The pipeline took it from there.
2. Twitter mentions. Reply to any thread with @andTadah now and the bot replied with a link to an audio version of that thread. Account linking happened through a low-tech verification flow. We issued a twelve-digit code, the user tweeted it publicly, the bot confirmed the tweet.
3. Web submission. A public endpoint on the site accepted a URL with no authentication, useful for people who wanted to try Tadah before signing up. Same pipeline behind the scenes.

Capabilities the work demonstrated

This was a production agentic system before the word “agentic” had settled into industry vocabulary. It is worth itemizing what the work actually proved we could build, because each piece transfers cleanly to the engagements our clients now ask us into.

• Multi-front-door ingestion. Email, social mention, and public web endpoint converging on one orchestrated pipeline. The same pattern sits inside half of the enterprise agents we ship today.
• Production audio synthesis. TTS provider abstraction, multi-track ffmpeg mixing, signed-URL delivery from object storage. Non-trivial systems work that people forget is hard until the audio sounds wrong.
• Token-aware LLM orchestration. Recursive chunked summarization for content that exceeded the model window. Obsolete now that windows are large. At the time, the difference between handling a three-hundred-word blog post and a twelve-thousand-word longform piece without truncation.
• Low-tech account verification. The tweet-the-code flow was an early reminder that the right engineering answer is often the right UX answer. Variants of the pattern show up in client work whenever OAuth feels like overkill.

What we learned

Tadah did not become the company we set out to build. The retrospective is more valuable than the product was.

The first lesson was about scope discipline. The vision was format conversion, language translation, and complexity simplification across every input source we could connect to. We shipped one slice of one axis (article to English audio, lightly summarized) and that slice was hard enough on its own. The translation and complexity-adjustment axes were where the real user pain lived, and they were the parts the model layer of 2023 was least ready to deliver on. Three thin verticals at once meant none of them got good enough to make a habit out of.

The second lesson was about evaluation infrastructure. We had no automated way to tell whether a summary was good. Listeners told us. That worked for a hundred users and would have been fatal at ten thousand. Every agent system we deploy now ships with evaluation harnesses on day one. Tadah is why.

The third lesson was about distribution. We assumed people would forward articles to a robot. Some did. Not enough. The Twitter mention flow was the sticky surface, and then Twitter changed its API pricing in the middle of the build and the bot economics inverted overnight. Building on someone else’s social graph is a posture, not a strategy. Forward-deployed work with named enterprise clients turns out to be a better fit for what we do.

The fourth lesson was about who we are. Tadah taught us we are better at shipping the agent inside someone else’s organization than at acquiring consumer users to one we own. The people who built Tadah now embed in client teams under the IdeaSparks banner. The discipline carried over. The product did not need to.