Punch · Founder's office
Founder's office at Punch
Two and a half years, two pretty different jobs under one title. For the first eighteen months I ran creator acquisition and rebuilt the referral program. For the last six I replaced most of that function with software I wrote myself.
What actually happened
I came in to run creator acquisition and spent eighteen months doing it end to end: sourcing, screening, deal structuring, and the call on where the next rupee of spend went. Then the jobs started arriving faster than a small team could absorb them, and Punch genuinely could not hire its way out of any of it.
So I started building the tools myself. I think the interesting part isn't the count, it's that four of those five systems replaced work I used to do by hand or paid someone else for. I made my own function smaller. That's an uncomfortable thing to put on a portfolio page, and it's true.
The creator program, 2024 to 2025
An eight-figure rupee creator-acquisition program at a discount brokerage, run inside real regulatory constraints. At its peak the channel accounted for roughly 46% of all company orders and brokerage in a single month, at a blended CAC about 44% below the platform's non-flagship acquisition cost.
The part I'd point at is the three product integrations. Not ad reads: each creator's own tool, built natively into the trading app and negotiated one at a time. H-Level put a well-known trader's daily levels onto any user's chart, LTP Calculator and PSBB shipped a creator's proprietary indicator as a native feature. Two ran on brokerage share plus a per-account payout, one on brokerage share alone, and all three returned differently: one drove the most volume and activations, one had the cleanest funnel, one retained best once a user actually converted.
One of those launches generated demand so far past internal projection that account-opening infrastructure, untested at that scale, broke under it. About 5,000 signup attempts in 24 hours. That's the second time in my career a creator campaign has taken down something I worked on. The first was smallcase, in 2019.
The referral rebuild
The industry standard is a flat "refer and earn brokerage credits", and it is standard because it is easy to approve, not because it works. I replaced it with a gamified mechanic built on a specific behavioural read: active options traders and gamblers share psychological wiring, so a shot at an outsized reward beats a small guaranteed one.
The design work was mostly in the constraints, not the mechanic. SEBI and NSE rules on how rewards can be structured are real and unforgiving, so the model had to work with no differentiated reward slabs and the incentive capped at the referred user's first trade. It lifted visit-to-account-opening conversion about 5% over the standard model.
Five production systems, 2026
I am not an engineer. I was a growth person with a budget and a set of jobs nobody was going to get hired to do, and the first tool worked well enough that I kept going.
Punch had no SEO function at all. No agency, no freelancer, no tooling, nothing to inherit. I built it and then ran it on under four hours a week. Organic traffic grew roughly 3x over six months.
That 3x is not all mine. A desktop product launch ran concurrently and my honest split is about 50/50 between the two. I have never claimed the full delta and I'm not going to start on a portfolio site.
The search half did what it was supposed to: roughly 3x organic in six months, from nothing, on under four hours a week. The GEO half, getting Punch cited by LLMs rather than ranked by Google, produced 0 out of 20 across five queries measured against four different models, and it stayed at 0 for four straight months. I published both numbers internally as they were. I didn't reframe the second one as an early-stage learning. I think knowing the channel was dead while it was still cheap to know is worth more than the 3x.
I evaluated two PR agencies, at ₹5.88L a quarter and ₹1.75L a month, each also charging a commission on top of media spend. I recommended against both and built the function in-house instead. It shipped 9+ live placements for about ₹4.25L in pure media spend, no retainer, no commission layer.
Stated plainly because it matters: every one of those placements is paid. None of it is earned media. What the system replaced was the agency's markup and coordination layer, not the fundamental economics of Indian tech PR.
A Claude-orchestrated video pipeline for Punch's YouTube channel, wiring together DaVinci Resolve, WhisperX, ffmpeg and Remotion across twelve stages. It cut a long-form video from roughly three to four weeks of manual production down to about a week, including my own review time. Three long-form masters plus reels were rendered and director-approved.
The design decision I'd defend hardest: Claude never emits a timestamp. A deterministic resolver converts spoken-phrase anchors into frame numbers. That denies the model the ability to produce that whole class of hallucination. Asking it to be careful is not a control. The QA gate refuses to run without fresh evidence, it doesn't warn and continue.
A creator-discovery and screening system where I designed the scoring logic myself and then personally reviewed all 225 discovered channels to calibrate what it should keep and what it should reject. The agent ran the discovery and the code, the judgement about what a good channel looks like was mine, and I don't think that division is worth blurring.
For scale: a fully manual pass in 2024 got through about 130 creators in one and a half to two weeks. I also used it to reconstruct Punch's 2024-25 creator-marketing data across sixteen source spreadsheets, which surfaced nine cross-sheet contradictions in numbers the company had been treating as settled.
My first classifier reported 99.6% accuracy. When I built a proper evaluation against a reject-all baseline, the real number was 73.1%, which is worse than the 85.8% you'd get by rejecting every channel unseen. I kept the measured number, wrote down why the first one was meaningless, and rebuilt it. The system is validated and ready to scale but has not yet been run at scale.
Sequence design, segmentation, deliverability engineering and the tests, all mine. It is also the only work I did at Punch with a clean funnel attached to it, so the experiments are worth walking through properly.
The onboarding sequence. Five emails. The first went to 761 contacts and opened at 29% with 1.8% clicks, by the fifth it was 619 sends, 10% opens and no clicks at all. That decay is the result. The sequence was too long, and the numbers said so before anyone had to argue about it.
The segmentation. The target segment came out at 1,908 contacts, derived from a 6,387-person Mixpanel cohort minus 13,464 registrants after I found the two groups overlapped about 70%. Skip the overlap check and a big chunk of that list gets mailed twice. That's how you burn a domain, not just a campaign.
The A/B test. Concluded 30 April 2026 with no winner: control 1,497 at 29% opens, variant 660 at 29%. The interesting number there is not 29%, it is 1,497 against 660. Loops' own native 50/50 split tool had silently delivered a 69/31 split. I caught it in the send counts and flagged the vendor's tool as broken. If I'd trusted it I'd have written up a result that was noise either way.
Deliverability, and one thing I refused to send. SPF, DKIM and DMARC alignment, RFC 8058 one-click unsubscribe, Google Postmaster spam rate held under 0.3%. There was a backlog of roughly 22,000 funded-but-never-traded users that I was asked about and did not blast. The deliverability damage would have outlasted anything it returned.
Loops is dashboard-only, with no API. Every send here was manual, so this is the one piece of 2026 work with no repository behind it, and the volume is small next to what an automated stack would push. The tests are honest because they were small, not in spite of it.
What I'd take to the next one
- Build the measurement first. Both of the numbers I'm proudest of here, the 0/20 and the 73.1%, are numbers that told me something wasn't working. Neither would exist if I'd built the thing and then looked for a metric to justify it.
- Structure decides the return more than the creator does. Three integrations, three different deal structures, three different shapes of result: most volume, cleanest funnel, best retention. Same channel, same product, same period.
- Constrain the model in code, not in the prompt. Arnold and Forge arrived at the same instinct months apart and independently: identify the class of output the model must never generate, then architecturally deny it the ability. Asking it to be careful is not a control.
- The agency question is usually a markup question. Two quotes, a quarter of the cost, same placements. What I could not replace was earned media, and I stopped pretending otherwise early.