# Parth Ahir

> I build agents that work inside real operations and earn their autonomy one approved action at a time. Building Aera and putting it to work inside Kalamandir Jewellers. Previously Toglo, Captain AI, Kalyxa, and Boeing.

- Location: San Francisco, CA
- Email: p.ahiir01@gmail.com
- GitHub: https://github.com/parth012001
- Education: Pennsylvania State University, B.S. Computer Science, 2020 to 2024

## What I believe

- 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.
- Trust is a property of an action shape, not of an agent. Earned on evidence, revoked on the first bad run, never widened for convenience.
- An eval that cannot tell a fix from model variance is not an eval. Take two baselines before changing anything; publish the negative result when that is what you found.

## Now

- Building Aera (2026 to present), an operations brain for multi-unit operators: seven to nine disconnected systems in, one daily brief plus approve-from-chat actions out, on WhatsApp. The current project builds it into an autonomous org for Kalamandir Jewellers, a jewellery house founded in 1986 that did around Rs 5,850 crore (about $700M) in FY25, starting on the vendor and procurement side: agents on top of existing systems that draft and reconcile purchase orders, handle vendor communication, and route exceptions to a human. Early deployment on the group's real data.

## Work

- Carrier Desk (public, https://github.com/parth012001/carrier-desk): AI carrier sales rep for freight brokerage. Negotiates inside a rate ceiling the model never sees; identity enforced in the tool layer. 610 offline tests, 6 adversarial personas, 28 dated decisions. The eval caught the model arguing itself into double-brokering; the tool layer stopped it. Also shipped as a private voice deployment against a legacy TCP TMS: 43 real calls, 9 bookings, 0 without an identity check.
- Toglo, LMS delivery (private, 2026): took an HVAC franchise from first discovery conversation to a live LTI 1.3 integration inside their production Moodle in under 90 days. 300+ tests, a 19-check live verification harness, two security findings caught before launch. Go-live 21 July 2026.
- Toglo, FieldMate (private, 2026): voice and camera coaching agent for HVAC technicians. Deterministic gate (frozen negation floor AND a narrow LLM judge, fails closed). Intent benchmark 35/45 to 44/45. Shipped in English and Russian.
- Greenlight (public, https://github.com/parth012001/greenlight): approval-gated IT agent. Policy engine, approval queue, hash-chained audit log, trust earned per action shape and revoked on one bad run, 50-action counterfactual replay before any rule widens.
- Meridian OS (public, https://github.com/parth012001/meridian-os): an autonomous organization for a distributor under a human-written charter. 82 tests, 6 graded trials including 3 red-team traps and a 26-order load run.
- Memory for an AI companion (design study, 2026): a memory architecture built on threads rather than facts, with a precomputed stable belief block, validity windows for facts that stop being true, and a human in authority over what is marked closed. 17-page design, grounded in a survey of six memory systems and the cognitive science of human memory.
- Captain AI (public, https://github.com/parth012001/capAI, 2025 to 2026): a chief of staff that sends the email and books the meeting behind an approval gate. 140 beta users, 90% action acceptance.
- Kalyxa (2024 to 2025): AI styling marketplace, web and iOS, Stripe Connect, team of 4, $150K seed commitment. Wound down.
- Boeing UAV (intern, 2023): real-time object detection, YOLOv5 at 95% and 30+ FPS on Jetson; ROS2 navigation, 85% lower simulated collision risk.
- Merged an E2B sandboxed code-execution MCP server into Klavis (YC X25): https://github.com/parth012001/e2b-mcp-server

## How I work

Twice a founder. Discovery with the customer, ship into the systems they already run, own the outcome when it breaks. AI coding agents do most of the typing; the architecture, boundaries, test bar, reviews, and decision log are mine. Boring primitives where the stakes are: integer cents, identity from the session and never from the model, default-closed, first-match-wins, append-only.

## Writing

- Policy belongs in the tool layer, not the prompt: /writing/policy-in-the-tool-layer
- The eval that could not tell a fix from variance: /writing/eval-fix-vs-variance
- Making an LLM safe enough to stand next to dangerous work: /writing/safe-next-to-dangerous-work