Pass 01 · Earth observation · 2019 → 2026
Founder & CEO of Sibkade. Now building Barayand.
Startup × Tech × Psychology. Six years in, still early.
Scene 01 · Upstream
The first draft of Sibkade.
Founder · 2019–2020
IranSpoti was a small gift-card business that lived entirely on Instagram. It lasted about a year, and that year was a year of trial and error: I taught myself WordPress, ran my first digital ads with people in the gift-card space, and learned how much a real website, search, and Google Ads mattered compared to a feed.
It never became a company. It became the plan for one. Almost everything Sibkade got right in its first year, IranSpoti got wrong first.
Scene 02 · Main stem
Building a gift-card business around customer experience.
Founder & CEO · 2020–present
Sibkade is a digital gift-card and subscription business serving customers in Iran. I founded it in 2020 and have led it from the first website and the first ads through building a team and developing the systems behind the business.
Gauge station S1
In 2020 I built the website and handled everything myself. I ran the Google Ads that brought in the first customers while managing the day-to-day operation. That gave me a close view of the whole journey, from the first ad someone saw to the questions they asked after buying.
As orders passed ten thousand a year, I hired and trained a team. My responsibilities grew with it, and I stayed in the decisions about the product and the customer’s experience.
Gauge station S2
We were the first in our market to deliver gift cards instantly, so customers receive their codes without waiting for someone to process each order by hand. If a paid instant-delivery order misses its window, the fee is refunded in full. Loyal customers get a VIP tier with their own support line.
The result is the growth channel we never paid for. Nearly one in two customers come back, and three in ten order three times or more. Most of the rest arrive because someone they trust recommended us. That is the standard I wanted: an experience people are comfortable putting their own name behind.
Scene 03 · First distributary
Could software notice the patterns in how you think?
Founder · 2023–2025
HelpFinity started from a question I kept running into with my background in psychology: people can name cognitive distortions in other people’s thinking, and almost never in their own. The idea was an app that reads what you write and points out the CBT patterns inside it, with journaling, breathing exercises, and short audio alongside.
The timing mattered. OpenAI had just opened its API to system prompts and then launched the first Assistants API, where you could give a model a role, a description, and tools. It was one of the earliest ways to build something like an agent, and combining it with a fixed taxonomy of thinking patterns was the first product I designed around a language model. I took it as far as a prototype and a plan; it did not become a company.
Coda, 2026. As a tribute to that idea, I rebuilt its core as a small Persian Telegram bot called Mind Mirror. You write the thought that bothered you. It names up to three patterns, quoting your own words, offers a more balanced reading, and asks one question back. Over time it shows which patterns repeat and where. It is deliberately not therapy: no diagnosis, a crisis screen that runs before anything reaches the model, and hard limits on what it will discuss. Fifty people are using it in a private launch. This is a side project, not a venture, and I keep it that way.
Scene 04 · Main stem, 2026
Gauge station S3
In 2026, after six years inside the business, I designed and built the core systems myself: twelve production systems in eleven weeks, about 41,000 lines of code and 15,000 lines of tests, all live.
Each one answers a practical question a customer asks. Where is my code? Automated code fulfillment. How much do I still need to pay? Wallet payments that show the split clearly. How do I get back into my account? Phone-first login and passkeys, after the data showed almost no customer ever used a password. When can I expect a reply? A support desk inside the customer’s account, with an honest estimate.
Behind fulfillment, I also built the tools for tracking stock, spotting shortages, and allocating codes when inventory is replenished. These are the less visible parts of making a delivery promise work.
Gauge station S4
The helpdesk is built around how our agents actually work: customer history, internal notes, follow-up reminders, mentions, and feedback on individual replies. The aim was enough context to continue any conversation, and a way to review the quality of support without reading every ticket.
I integrated AI into that workflow. It helps draft and refine replies using product-specific guidance, and turns recurring questions into proposed knowledge-base entries for managers to review. Staff see suggestions before anything is sent. A separately controlled option can answer automatically, only from approved knowledge.
Gauge station S5
I built the customer-intelligence and reporting tools that bring purchase history, customer segments, churn risk, and product-level profit into the daily operation. I also built first-party tracking for sign-in, checkout, and payment, because off-the-shelf analytics recorded zero purchases on our stack. Now those parts of the journey can be measured, not guessed.
Scene 05 · Braided reach
One question, a panel of frontier models, one answer you can measure.
Founder · 2026–present · barayand.io
Barayand is the Persian word for a resultant: the single vector you get when several forces are added together. That is the product. Instead of trusting one model, you send a question to a panel of the world’s leading models at once. A judge model reads all the answers and extracts where they agree, where they diverge, what only one of them noticed, and what all of them missed. A synthesizer then writes one grounded answer, with an agreement score attached, so you can see how much to trust it.
The idea came from a simple observation: every model is confidently wrong sometimes, and no single one is right all the time. Disagreement between models is not noise. It is signal, and Barayand turns it into a product.
I started building it in June 2026 and shipped it the same summer as a Persian-first platform: chat with any model, image generation, long-term memory, web and deep research, gift subscriptions, local pricing, and phone login. Two assistants sit on top of it. Hambonyan turns a business idea into a workspace, runs the playbooks for customers, market, competitors, and unit economics, and ends with a board review where a panel of models argues against the idea before it gets a verdict. Hamfekr guides a research project from question to defense, keeping the writing in the author’s hands.
Alongside the product, I published an open, reproducible benchmark for Persian: 300 questions across 10 axes, deterministic scoring with bootstrap confidence intervals, and reference results from the leading models. The Persian-speaking world needed a shared yardstick, so I built one.
I designed and built Barayand alone, from the brand mark to the billing.
Scene 06 · Delta front
Things I build on weekends, for the fun of it. None of these are companies. All of them are finished.
A personal archive that eats everything you save and gives it back when it matters. Reels, podcasts, screenshots, book pages, voice notes: it transcribes, summarizes, and files them, then resurfaces the right one on your phone when you are likely to want it. Built on the bet that people do not pay to search their own saves; they pay to be reminded. I use it every day.
A Persian Telegram agent for students applying to graduate programs abroad. It interviews you, researches real programs on the live web, reads the official admissions pages, and writes a tailored CV and statement of purpose. It never submits anything for you. My first end-to-end agent with tools, memory, and a budget.
A clock that refuses to behave like one. A planet hangs in the dark, the time floats above it, and the day runs at that world’s real length: a Martian day is 37 minutes longer than ours. Tilt the phone and the planet lags behind, as if it had mass. No account, no notifications, no streaks. Time, kept at a distance.
On Mars now --:--:-- MTC · Sol -----
A single page that turns your scrollbar into 3.8 billion years. You start as one glowing dot. As you scroll, the branches of the tree of life merge into yours, one by one, until everything converges on the last universal common ancestor. Then the camera pulls back and you are one lit tip among millions. No libraries, no build step.
A benchmark for how well language models actually handle Persian: 300 items across 10 tracks, from reading and culture to hard math and strict instruction following, with deterministic scoring and bootstrap confidence intervals. I built it because I wanted a Persian yardstick to exist, and nothing did. The result I did not expect: on strict instruction following, the biggest and most expensive model lost to a smaller one.
I learned Claude Code by building the things I needed from it. Venture is a business copilot: market, competitors, customer discovery, unit economics, go-to-market, all feeding one shared workspace and ending in a board review. Idea Validator turns the idea chapter of Anthropic’s Founder’s Playbook into a ten-stage pipeline with kill criteria written first, evidence graded by its source, and one wall it refuses to cross: it will not fake customer interviews. Liquid Glass teaches Claude Apple’s iOS 26 design language, with the API, the guidelines, and nine working screens.
Ground truth
I’m Parsa. I run Sibkade, a digital gift-card business I started in 2020, and I’m building Barayand, a multi-model AI platform for Persian speakers. I trained as a psychologist and write about AI in Persian most days.
Disagreement between models is signal. Barayand sends one question to a panel of frontier models and writes one answer, with the confidence attached.
Restraint builds trust. My psychology bot’s best feature is what it refuses to do. So is my clock app’s.
Measure it, or you don’t know it. There was no Persian benchmark, so I built one. The most expensive model came out worse on one track.
2019 IranSpoti · 2020 Sibkade · 2023 HelpFinity · 2026 Barayand, and the weekend projects
X @parsakzn · Telegram @parsa_notes · GitHub heyparsadev · LinkedIn · me@heyparsa.com
Reads Camus. Thinks Apple is the best company on earth.
In March 2023, a week after GPT-4, a thousand people signed a letter asking to pause training bigger models. I remember thinking they were killjoys. Three years later, I’m one of them.