Zinta

A studio our size lives or dies on a steady pipeline, but founder-led outbound is the first thing to collapse when client work gets busy — exactly when you need it most. The obvious fix, bulk cold email, is worse than doing nothing: it burns the sending domain, annoys good prospects and produces the template-shaped mail everyone deletes. We wanted the opposite: a small number of genuinely researched companies a day, each mail specific enough that it could not have been sent to anyone else, and a human deciding every single send.

• Wrote the agent as a specification, not a script. Zinta’s behaviour lives in a single reviewable markdown file covering targeting, research, pricing, mail craft and compliance — so changing how she works is an edit, not a deploy.
• Grounded the mail rules in published cold-outreach research rather than instinct: 50 to 125 word bodies, 2 to 5 word subjects, a trigger opener carrying a fact true only of that company, one soft interest-based CTA, and the first line treated as inbox preview text because in plain-text mail that is exactly what it is.
• Made honesty a hard constraint. She never guesses an email address or constructs a LinkedIn URL from a name — a blank field is required where a real one could not be found, because a fabricated contact is worse than no contact.
• Split the work by trust: the model researches and writes, while a small deterministic Node bridge does the dedupe, merge and persistence. The parts that must never be creative are code.
• Priced to the market rather than a flat rate, quoting in the prospect’s local currency across ten of them, with the INR equivalent shown internally so deal size reads at a glance from Bangalore.
• Built the review cockpit as the product: an accordion queue sorted by fit score, an email preview rendered the way the recipient will see it, and every reach-out channel — contact page, company and personal LinkedIn — on the card for follow-up day.
• Designed the failure modes first. A run aborts rather than write a partial batch over good data, the console cannot overwrite work the agent added while a tab sat open, and an unreachable store fails in seconds instead of after an hour of wasted research.

• A daily run that surfaces up to 25 client prospects and 5 agency delivery partners, each with a real, cited buying signal rather than a scraped list.
• Mails that hold to the spec measurably — recent batches sit at 93 to 116 words, zero em-dashes, an opt-out line on every one, and quotes in nine different currencies priced to their market.
• Contact discovery that reports its own gaps: verified addresses are found for roughly half a batch and the rest are left deliberately blank, so the queue never hides a guess as a fact.
• A human approval step that has never been bypassed. Zinta drafts, a person sends.
• Proof of the thing we tell clients: we do not just build AI agents, we run one in production every day and live with its output.