so after a super long period at my first company, leading ai development for a revolutionizing intelligent monitoring system for pemra pakistan that would help them choose what to censor better, there was a lot of unnecessary business delays and government project shit going on (you know how government infra works in pakistan) that caused us not to scale and grow much.
after a year of trying my best to scale the team, create a system, grow the project, and hunt for any personal growth opportunities for myself, i didn't really find a way forward. hit a roadblock. same routine for almost a year and same loop, wake up, go to office, coffee, table tennis, some maintenance shit, overseeing interpersonal issues and then dinner out, would just end the day. so yeah, no growth really.
and while i did some courses and stuff and also did the aws ccp, i finally decided that i had to leave. there was growth in terms of money but i needed something more than that in just the second year of my career. i wanted to work on sota tech and stuff and was hungry to learn.
gen ai freelancing chapter
gen ai had also just come in the scene around that time and i had been working with dall-e 1 and gpt-2, 3 way before chatgpt and llms were cool, so i was like let's make use of that experience.
i did have okayish product development experience along with team management, people management and stakeholder management experience (i dealt with pemra execs. i can deal with anyone. bring it on), but i didn't know one thing that i lacked and that was business and operations knowledge.
so after leaving the first company, i started offering gen ai freelance and consulting services to products/startups in the ai space, and boy what an experience it was. my first client didn't even pay me as i didn't yet know how payments worked on upwork (figured it out by my second gig though).
i got to meet a variety of clients, some very smart, some very ambitious to say the least (i still remember the arab guy on fiverr that wanted me to build his own chatgpt, just me and him). i met some very good people and did get to learn a lot about what goes on in the mind of an entrepreneur, how to approach a problem effectively, how to integrate tech in a physical business and other valuable stuff.
freelancing also teaches you ownership, thinking outside the box, but your code quality would mostly suck, and you keep adding to tech debt most of the time.
during this journey i worked with amazing groups, a research team from MIT where I helped implement an avsr paper they were working on, a health tech startup by the name of clinomic, revolutionizing icu care in germany.
i got to learn a lot about infra, production, scale, research, applications of ai, commodification of tech, what's possible and stuff.
the yahoo of gen ai
then i joined an nlp talk-to-your-data chatbot company which was still practicing statistical models for their tech and i suggested my team lead and ceo if we should explore probabilistic ways of doing this (gpt-3.5 was out by then), and he said something so absurd that i had to resign exactly at that point.
they did have good clients, accenture, royal bank of canada, but the tech was outdated and worst part, they were ok with that, and did not want to evolve.
layoffs, breaks & eduard
after that, i joined a services company based in las vegas, and they were pretty cool, but after like 45 days i found out that oakland airport and lvpd had pulled the contract from them and kenwood hardware didn't come through. to cut it short, they just laid off the ai team that they formed 45 fucking days ago, all three of us, the team which took them 2 months to hire.
i took a little break after being laid off, bought a mountain bike, started cycling in the city, something i used to do a lot as a kid. and around that time i started working with eduard from germany, who is now a great friend of mine.
we tried really hard to get mono-channel diarization working in real-time, but the research and tech available at the time was immature and not accurate enough, and it was, in its roots, a very difficult problem to solve as well. but we were ambitious and did try a very long time before deciding to shelf the thing.
it was very important to eduard as he was building it for his mother, who happened to have a hearing impairment. so i was a bit bummed out that it didn't work out.
train a nnet on 5th gen thinkpad, can you?
a little while after that, i was onboarded by another services company, and they tasked me with setting up an ai team in-house. i was hired as a consultant, and had to manage hiring, clients, business, and sometimes who should get the new laptop and who should be fired (something i did not want to do, at all).
the honeymoon phase was great, and i did build a very good team by setting the standards very high, but the senior leadership lacked vision and understanding of how the ai field worked and didn't really invest like they promised, they would literally want the team to do deep learning research on 5th gen lenovos.
even after all that, we ended up generating very good business and growing the team to 7 people, and actually delivering and scaling ai to some of the clients. i personally believe we could have done wayy moreee than that. as soon as the contract ended, i left the company and never looked back.
the first founder itch
during my time at this last company i wanted to build something of my own, owning it end to end. i wanted to build a product, a tech product, and make a business out of it. i had done it for clients and companies before, and now i wanted to do it myself.
fast forward a few months, i met aaron and mostafa. aaron worked for the us fed govt at the time while mostafa was on the chrome team at google. aaron and i were really big fans of music and i have been doing music as well when i was young (spoiler: i used to sing while in school and college).
we wanted to do gen ai for music. we built a very small team and started experimenting with design, ai backend for lyrics, composition, vocals and synthesis of all three.
but after some 6 months of research and trials with open source work, the available tech was super immature, inaccurate, and the operating cost was too much for us at the time (around a thousand dollars if we were to serve a single user for a month).
progress was also slow since we were in different time zones (gulf, oceania, north america) and since we couldn't build or train a foundational model for the problem because of not enough resources at the time, we worked with what was available open source only, we tried audiocraft, jukebox etc., but didn't really get anywhere, and ended up shelving musafy.
then suno ai also just announced their demo and had come into the scene so we decided it's over because these five guys from cambridge were so fucking good.
pivot mania
aaron now wanted to build something for students from his college, the idea was an ai-assisted research writing platform that would assist with tones, drafts, citations, references and research all within the same application (sort of like an all-in-one research writing platform, think latex but ai powered and with outside knowledge connectors).
but after a bit of research there wasn't much value proposition that we could offer and both rytr and grammarly would offer nearly every feature we could think of, so no point re-inventing the wheel. we didn't really go beyond the design and use case study phase and pivoted yet again.
also, citations and sourcing through llms required a lot of effort, and llm hallucination was still a huge challenge. even if we were to bypass that, a team like grammarly could just integrate it and no one would use us for just one feature, we would have to build infra like them to gain any traction and to scale.
after a few months, we met this guy from lexisnexis and showed him what we had been building for a year, and he was impressed with some of the implementation. he wanted us to do the same research and writing platform but for law.
so we pivoted again and umiiro was born.
umiiro offered legal research, drafting, and a few other tools to help attorneys and people in law. i took it very seriously this time and worked super hard, i scraped the whole us constitution, all the federal laws, everything legal available online and fine-tuned llama.
i also tried some domain specific llms (saul lm) but they would suck and not perform very well, even with tweaks and tuning. also, fine-tuning is expensive, and we couldn't afford to do it the third time.
this time the pitch was strong and the product not too bad, but the presentation didn't really go very well, and we shut down after a while because what we could build with the resources we had was not convincing enough for them.
i concluded that we lacked conviction, agency, and there were a lot of delays due to difference in time zones, progress was very slow, and all three of us were balancing it with full-time jobs.
maybe some of the tech was not there for what we were trying to build and for some things, no value proposition, not enough for a customer to pay. and sometimes we would obsess a lot over design and minor stuff and wouldn't actually build the thing out soon enough.
building for europe
after the three musketeers period, i met philippe, with whom i was working on a project for a while and one day he asked me if i wanted to work on a product together, and without a second thought i said yes because i loved the idea, the value proposition, and the problem it solved
so i now started working on dynomo, which is a dynamic pricing solution for SMEs and restaurants in europe. this time i went in even harder.
we got the iot from south korea, attended a few conferences to figure out all the physical tech we needed, researched a lot of deep learning-based tech to get the best algos in the system, managed to bypass some GDPR issues regarding collecting customer data with proper compliance, pitched to some investors, went to YC (didn't get in), pitched to restaurant founders, but just couldn't get it live.
there was a lot of installation and integration overhead, and some of the tech was overengineered (lol), the cost-to-value ratio was a bit messed up.
this one hurt a lot, as i overdid it this time, perfecting everything, late nights, writing drivers for iot myself (yeah I learned some c++), figuring a lot of shit out to get it working, but even after all that, not getting anywhere.
i figured at the end that what i was working on and what the customer on ground needed was a bit different, and that you should always try to find simple solutions to the most complex of problems.
the pause, reflection, and going again
after all this effort and pursuit, i was a bit heartbroken. so i took some time off from all the product fuss.
i did pursue one or two more ideas, one a sports management app with a friend from my football team, and we did focus on sales beforehand this time, but that didn't go very far as our sales guy on ground could not deliver and i couldn't at the time source another guy in the market that i was building for.
and then another futsal analytics app, as i have been playing the game for 11 years myself, but couldn't sell it to the owners or fit a proper business model for this one in the local pakistani market as the whole football ecosystem is still developing and the vision within stadium or arena founders is yet to be developed. they do not see the value yet.
while there wasn't much success in terms of traction or sales or money, it was a great learning experience, and it was my start within the entrepreneurial world.
i took a lot of risks, didn't follow the normal career trajectory, did my best with the knowledge that i had, but learned that my best was not enough (so far).
so i have been working on my best ever since and will go again, start something again very soon.
with more grit, more belief, more agency, and with all the experience i have gained from failing over and over again.