About

I grew up in Surat, India, and moved to the United States for Penn State, where I finished a computer science degree in 2024. I live in San Francisco.

Boeing was the first time I shipped machine learning that had to work outside a notebook. In the fall of 2023 I was a software engineering intern on a UAV program: real-time object detection with YOLOv5 tuned to run at 30-plus frames per second on NVIDIA Jetson hardware in field conditions, and ROS2 navigation and obstacle avoidance that cut simulated collision risk by 85 percent. Eight people, cross-functional, weekly stakeholder reviews. It taught me that the model is the easy part and the deployment target is the whole problem.

Founding came next, three times. Kalyxa (2024 to 2025) was an AI styling marketplace: two products, web and iOS, Stripe Connect payouts, a team of four, a seed commitment from a pre-accelerator. We wound it down in June 2025. It taught me to pick a wedge and de-risk the hard part before building anything else. Captain AI (2025 to 2026) was a chief of staff that took real actions behind an approval gate; 140 people used it and accepted 90 percent of what it proposed. It taught me that the approval gate is the product, not a compromise on it. Toglo (2026) was an AI training platform for skilled trades, co-founded with an ex-Retool product leader; I took our first customer, an HVAC franchise, from discovery to a live integration inside their LMS in under 90 days. It taught me that selling into an operator means deploying into whatever they already run.

Now I am building Aera, an operations brain for multi-unit operators who run seven to nine disconnected systems and get one daily brief, with approve-from-chat actions, on WhatsApp. The project I am on is putting it to work inside Kalamandir Jewellers, a roughly ₹5,850 crore Indian jewellery house, starting on the vendor and procurement side.

How I work

• Discovery first. I sit with the people who do the work until the real process, with its exception paths, is written down.
• Ship into what already exists. The LMS, the TMS, the ERP. Never force a migration.
• The model proposes; a deterministic layer decides. Identity comes from the session, never from the model. Default closed. First match wins. Append-only logs.
• Two baselines before any change. If the eval cannot tell a fix from variance, I say so.
• AI coding agents do most of my typing. The architecture, the boundaries, the test bar, the reviews, and the decision log are mine.
• I write the honest gaps into the document before a reader has to find them.

Things I am interested in

• AI for the physical world, the half that augments the person doing the job instead of replacing them with a robot: a model that sees what a technician sees, coaches them with their hands full, and earns trust because a wrong step has physical, sometimes irreversible consequences. Toglo put me onto this, and it is the open thread I keep pulling on.
• Agent autonomy that is earned per action shape rather than granted per agent.
• Operations software for the trades and for multi-unit operators: HVAC, freight, distribution, hospitality.
• Evals as evidence rather than as a checkmark, and what it costs to measure an effect size honestly.
• Deploying into legacy systems: LTI, raw TCP, Moodle, ERPs. The unglamorous integration is usually why there is a customer.
• Typography and quiet websites. This one is text first on purpose.

Timeline

p.ahiir01@gmail.com · GitHub · LinkedIn