Parth Ahir

I build agents that work inside real operations and earn their autonomy one approved action at a time. Right now I'm building Aera, an operations brain for multi-unit operators, and putting it to work inside Kalamandir Jewellers, a roughly ₹5,850 crore Indian jewellery house. Before that I co-founded Toglo, where I took an HVAC franchise from first conversation to a live integration inside their LMS in under 90 days; built and monetized Captain AI, a chief of staff that 140 people used and whose proposed actions they accepted 90 percent of the time; co-founded Kalyxa; and shipped real-time object detection for Boeing UAVs. Penn State computer science. San Francisco. I reply to every email: p.ahiir01@gmail.com.

Now

Building Aera, an operations brain for multi-unit operators. The project I'm on now puts it inside Kalamandir Jewellers, a roughly ₹5,850 crore jewellery house, starting on the vendor side: agents that sit on top of the systems they already run and take over the procurement and purchase-order work, with a human approving anything consequential. More on the now page.

Selected work

  • Carrier Desk · An AI carrier sales rep that negotiates inside a rate ceiling it cannot see. The eval caught the model arguing itself into double-brokering a load; the tool layer stopped it. 610 offline tests, six adversarial personas, and a published negative result. Also shipped as a voice deployment that placed 43 real calls and booked nine loads with zero identity-check misses.
  • Toglo, LMS delivery · Discovery to a live LTI 1.3 integration inside an HVAC franchise's production Moodle in under 90 days. 300+ tests, a 19-check live harness, two security findings caught before launch.
  • Toglo, FieldMate · A voice and camera coaching agent for technicians with their hands busy. A deterministic gate a too-agreeable model cannot open. Intent benchmark 35/45 to 44/45. Shipped in English and Russian.
  • Greenlight · An IT agent that acts when policy allows, asks a human when it doesn't, and earns autonomy per exact action shape, then loses it on one bad run. Hash-chained audit log, 50-action replay before any rule widens.
  • Meridian OS · An autonomous organization for a distributor, run under a human-written charter. 82 tests, six graded trials including three red-team traps and a 26-order load run.
  • Memory for an AI companion · A memory architecture built on threads rather than facts, so the system remembers you between conversations and never repeats a closed chapter. A design study grounded in a survey of six memory systems and the cognitive science of how human memory works.
  • Captain AI · A chief of staff that sends the email and books the meeting instead of drafting. 140 beta users, 90 percent action acceptance, paid for by account executives.

Everything I have shipped, including what I shut down.

Three things I hold

Each with a repo behind it.

  1. Policy belongs in the tool layer, not the prompt. A prompt rule is a suggestion the model weighs against whatever the user just said. A tool rule is arithmetic. → Carrier Desk, Greenlight
  2. Trust is a property of an action shape, not of an agent. It is earned on evidence, revoked on the first bad run, and never widened for convenience. → Greenlight, Meridian OS
  3. An eval that cannot tell a fix from model variance is not an eval. So I take two baselines before I change anything, and I publish the negative result when that is what I found. → Carrier Desk, FieldMate

Writing

How I work

Twice a founder. I do discovery with the customer, ship into the systems they already run, and own the outcome when it breaks. I use AI coding agents heavily; the architecture, the boundaries, the test bar, the reviews, and the decision log are mine, and every repo says so in its own words. Boring primitives where the stakes are: integer cents, identity from the session and never from the model, default-closed, first-match-wins, append-only.

Elsewhere

GitHub · LinkedIn · p.ahiir01@gmail.com · /about · /now · llms.txt