Adi Middha

Ever feel icky getting Venmo requested $7.43 for a coffee? Friends don't let friends track small debts down to the nearest cent. I love it when someone just tells me "I'll get you on the next one" or "I'll buy you a drink later". It's more fun, more personal, and guarantees that we’re going to hang again soon. Introducing IOU, a web app I just shipped, that tracks thing-based ledgers between friends: iou-app.vercel.app 🤝

My vision for the agent goes one step further than smarter predictions: better inputs - Automatically break down complex tasks into smaller, estimable parts - Create a list of common actions and build templates out of them (morning/nighttime routine) But that may be getting ahead of myself. For now, I’m plugging away at making this estimator GREAT. Dabbling in some AI eval frameworks and context eng strategies, will see if any of them are worth the squeeze

The current UX for my time calibration agent: Estimate → Get prediction Log actual → Learn patterns → Improve The better it learns you, the more you'll be able follow through on your personal schedule. Fewer days where everything looks reasonable on paper…but nothing gets done. Here’s what it looks like so far:

Have tweaked the algorithm in small ways since, but in general, this heuristic works decently well for adjusting to user patterns And most importantly for initial testing, I don't get wild predictions that would be immediately disregarded by an end user

But within the first few test estimations I noticed the algorithm was too aggressive. It applied the full bias score after just 1-2 tasks and had no upper limit on bias adjustments. I told Cursor that I wanted the adjustments to scale with the sample size and limit max bias, it generated the following edit: --------- confidence = total_tasks / 10.0            → if total_tasks < 5 0.5 + (total_tasks - 5)/10.0  →if 5 ≤ total_tasks < 10 1.0                            → if total_tasks ≥ 10 capped_bias = max(-0.3, min(0.5, raw_bias)) bias_factor = 1.0 + (capped_bias × confidence) total_factor = (category_factor × ambiguity_factor × bias_factor)^(1/3) total_factor = max(0.5, min(2.0, total_factor)) adjusted_minutes = base_minutes × total_factor

My starting architecture has 4 parts: 1. Storage → JSON persistence 2. Agent → OpenAI (gpt-4o-mini, cost efficient) for initial estimates, includes historical context 3. Learning → Bias algorithm that calibrates prediction 4. CLI → User interface, using Rich-based terminal UI so it looks pretty I paused on linking a real backend or moving to a more user-friendly web app. Wanted to get the core functionality working first.

Been using @WisprFlow for the last few months and it makes prompting way more fun Just sit there and blab for a few minutes and tell the LLM to format it into instructions Starting my new project on Cursor and a banger first prompt got me a working CLI prototype in a few mins

I realized recently I’m often too optimistic with time—especially on new tasks with no frame of reference. Turns out this is called time blindness. Figured I can’t be the only one, so I’m experimenting with an AI agent that learns and corrects my time estimation bias ⏳

My broad #vibecoding takeaways: 1) Start before you’re “ready.” Trying to build sub-optimally was weirdly a cheat code. I kept wanting to research “best AI tools / best prompts” first… which was probably just masked procrastination. Treating it like childish play (low stakes) is what sped up my learning. 2) AI dev tools aren’t “anyone can build” yet. The marketing for these tools implies zero barrier, but I still leaned on my coding experience: GitHub plumbing, understanding database changes, interpreting JS/TSX. Without that, it’s easy to get stuck or follow a misguided suggestion. 3) But for technical PMs: the time is now. If you can get a half-baked prototype live in a day, you can probably 10x how quickly you test ideas. Future prediction: the real valuable PM skill will soon shift to taste (what to build quickly) + customer understanding (what acquired feedback actually means). Overall: Lovable → Bolt → Vercel was an insanely effective pipeline. Now that I’ve stitched it together once, I'm pumped to make my future builds even better.

A few things I’m excited to try next to tighten this full-stack loop: - Figma: rework the UI after brushing up on design methodology - Claude Code: have it reason through my codebase + compare it to Bolt for feature builds - HCI throwback: revisit my grad notes on user interviews, screen recordings, and “time-to-click” type metrics.

I built IOU — a playful web app for tracking what you owe friends (in beers/rides/coffee) — in less than a day. Here's how I went from idea → prototype → real user feedback fast (Lovable → Bolt → Vercel)👇