# Arpit Arun Kumaar (arpitak)

> Arpit Arun Kumaar (arpitak) — Honours Computer Engineering student at the University of Waterloo working on Product (MCP, Plugins, Evals & Agentic UX) at Carta. Previously Applied AI & Venture at Axiom Partners VC, Silicon IP Design Engineer Intern at Low Power Futures, Software Engineer Intern at Intrepid Financial, Systems Engineer Intern at St Marys Cement, and AI/Robotics researcher at the University of Waterloo.

Hi! Welcome to my corner of the internet. I value thoughtful disagreements and am always curious to see things from other perspectives. I also lose track of time playing the guitar and sitting by the ocean.

## About

- **Name:** Arpit Arun Kumaar (arpitak)
- **Education:** Bachelor of Applied Sciences (BASc), Honours Computer Engineering Co-op at [University of Waterloo](https://uwaterloo.ca/) (Expected Graduation: March 2027; Awards: Term Distinction, President's Scholarship)
- **Current Role:** Product — MCP, Plugins, Evals & Agentic UX at [Carta](https://carta.com/) (Kitchener, ON, Sept 2026 – Present)
- **Contact:** [arpitarunkumaar@gmail.com](mailto:arpitarunkumaar@gmail.com)

## Education

- **University of Waterloo** — Bachelor of Applied Sciences (BASc), Honours Computer Engineering Co-op (Expected Graduation: March 2027)
- **Awards:** Term Distinction, President's Scholarship

## Experience

### [Carta](https://carta.com/) — Product — MCP, Plugins, Evals & Agentic UX (Kitchener, ON · Sept 2026 – Present)

- Built sentiment evaluator to grade user sessions from MCP/plugin telemetry, improving scoring alignment from 50% to 95%; surfaced unmet customer requests that directly shaped >6 new plugin features and informed targeted client re-engagement.
- Designed and built an agent-native experience prototype guiding investors through automated reporting setup for bespoke firm needs via interactive in-chat controls with MCP Apps and live report previews (won internal hackathon).
- Automated feedback filing to open PR loop: enabled agents to auto-file MCP bugs and feature requests on users' behalf and routed them via Airflow to an internal coding harness for triage and automated PR creation, addressing 100+ reports in 2 weeks.
- Quantified client-base AI tool adoption by count and ARR using AWS Bedrock to analyze client calls and inform GTM and integration priorities.
- Led sessions with VC firm clients demonstrating agentic workflows tailored to their AI familiarity, usage, and MCP/plugin adoption.
- Benchmarking open models to guide internal model selection, with planned GPU fine-tuning and autoresearch experiments.

### [Axiom Partners VC](https://axiompartners.vc/) — Applied AI & Venture (San Francisco, CA · Feb 2026 – Apr 2026)

- Led development of MCP aggregator server and skill workflows to expedite VC sourcing with agent-driven founder research and outreach flows, saving ~10 hours/month per partner.
- Developed eval framework for MCP server to trace agent failure patterns; iteratively refined tool schemas, descriptions, and scoring rubrics, reducing tool call error rates by 82% and improving LLM response quality from 65% to 91%.
- Built automated legal-ops workflow to pull context, fill, and validate investment paperwork, cutting $10k annually in direct costs.
- Synthesized ML research and competitive analysis into investment memos for the Managing Partner, informing evaluation of Seed–Series C opportunities across AI infrastructure and semiconductors.

### [Low Power Futures, Inc](https://www.venturelab.ca/portfolio/low-power-futures) — Silicon IP Design Engineer Intern (Markham, ON · May 2025 – Aug 2025)

- Extended EDA toolchain support after identifying missing Lattice and Intel integrations, accelerating prospect IP validation timelines.

### [Intrepid Financial](https://intrepidatg.com/) — Software Engineer Intern (San Francisco, CA · Jan 2024 – Jun 2024)

- Designed A/B tests across dashboard UX after identifying pain points in flows and onboarding, driving >10% reduction in support ticket volume.
- Engineered core creator analytics dashboards with TypeScript and TanStack, iterating weekly upon continuous customer feedback.

### [St Marys Cement](https://vcna.com/) — Systems Engineer Intern (St Marys, ON · Jan 2024 – May 2024)

- Diagnosed alert fatigue as root cause for 90% of missed fault actions through operator shadowing and interviews; led HMI UX redesign and devised priority-based alert architecture, achieving an 80% decrease in missed alarm rates and 70% visual noise reduction.
- Engineered deadband algorithm-based program to centralize instrument data for operations tracking and failure diagnosis across 3 sites.

## Research & Hardware / AI Projects

- **Human-centric AI Research — Conrad School of Entrepreneurship & Business, University of Waterloo** (May 2026 – Present): Investigating how LLM use influences human judgment under pressure, measuring willingness to edit AI-drafted messages and general confidence in AI responses; analyzed pilot findings to refine experiments testing how interface nudges affect both outcomes.
- **Research Assistant — [Vision & Image Processing (VIP) Lab, University of Waterloo](https://github.com/huggingface/lerobot)** (May 2025 – Aug 2025): Benchmarked performance of imitation learning vs. Vision-Language-Action (VLA) models on robotic pick-and-place tasks; fine-tuned SmolVLA policies for LeRobot SO-101, demonstrating stronger out-of-distribution generalization with up to 25% higher success rates on unseen object configurations.
- **[I-BERT Based LLM Hardware Architecture](https://arpitak.com/writing/ibert-fpga-impl)** (SystemVerilog, Vivado, Xilinx PYNQ Z1/Z2 FPGA, AXI-Stream): End-to-end quantized Transformer (IBERT) language model implemented in SystemVerilog RTL with 4x4 systolic arrays, cyclic memory bank address generators, pipelined Exp/GELU/LayerNorm, and FSM division, generating tokens natively on an FPGA.
- **[Energy-Aware Spiking Neural Network (SNN) on an FPGA](https://github.com/arpitarunkumaar/energy-aware-fpga-snn)** (SystemVerilog, FPGA, Spiking Neural Networks): Spiking Neural Network (SNN) implemented on an FPGA to predict collisions.
- **FP4/FP6 Sum-of-Products Hardware Unit** (SystemVerilog, Digital Hardware Design, Low-Precision Arithmetic): Low-precision floating-point (FP4/FP6) hardware sum-of-products compute unit.

## Skills

- **Technical:** Python, SQL, Swift, C, TypeScript, React, TanStack, SystemVerilog, Vivado, GCP, AWS Bedrock, Airflow, Supabase, Snowflake, Docker, Pandas, Linux, Agentic AI Patterns, Model Context Protocol (MCP), MCP Apps, Vision-Language-Action (VLA) Models / LeRobot
- **Product & Ops:** LLM Eval Design & Error Analysis, Agentic UX, Product Analytics, Experiment Design, A/B Testing, Customer & Market Research, AI Infrastructure & Semiconductor Venture Research

## Main Pages

- [Home](https://arpitak.com): Bio, current work, past experience, projects, and recent activity ([Markdown](https://arpitak.com/index.md))
- [Writing](https://arpitak.com/writing): Essays, technical hardware/AI write-ups, and personal reflections ([Markdown](https://arpitak.com/writing.md))
- [Reading List](https://arpitak.com/reading): Books recently read, to-read queue, and favorite newsletters ([Markdown](https://arpitak.com/reading.md))
- [Highlights](https://arpitak.com/highlights): Curated photo highlights from 2023 to 2026 ([Markdown](https://arpitak.com/highlights.md))

## Writing

- [The Songs We Keep](https://arpitak.com/writing/songs-we-keep) (2025-12-20, tags: personal): on music, neural nostalgia and EOY reflections ([Markdown](https://arpitak.com/writing/songs-we-keep.md))
- [An Intimacy Conundrum](https://arpitak.com/writing/intimacy-conundrum) (2025-11-02, tags: reflections): collection of thoughts and conversations from the past few months ([Markdown](https://arpitak.com/writing/intimacy-conundrum.md))
- [IBERT LLM Implementation in Hardware](https://arpitak.com/writing/ibert-fpga-impl) (2025-09-05, tags: fpga): Implementation of an LLM in hardware using SystemVerilog ([Markdown](https://arpitak.com/writing/ibert-fpga-impl.md))

## Machine-Readable Endpoints & Feeds

- [Full Site Corpus for LLMs (llms-full.txt)](https://arpitak.com/llms-full.txt): Complete markdown compilation of bio, all full-text articles, reading list, and highlights
- [RSS 2.0 Feed](https://arpitak.com/feed.xml): Full-text RSS feed of writing
- [Atom 1.0 Feed](https://arpitak.com/atom.xml): Full-text Atom feed of writing
- [JSON Feed 1.1](https://arpitak.com/feed.json): Structured JSON feed of writing
- [XML Sitemap](https://arpitak.com/sitemap.xml): Canonical URLs and last-modified timestamps
- [OpenAPI 3.1 Specification](https://arpitak.com/openapi.json): Public read-only JSON API specification
- [RFC 9727 API Catalog](https://arpitak.com/.well-known/api-catalog): Linkset catalog of APIs and documentation
- [AI Catalog](https://arpitak.com/.well-known/ai-catalog.json): Domain-level AI/agent discovery metadata
- [Profile JSON API](https://arpitak.com/api/profile): Structured JSON representation of Arpit Arun Kumaar's profile
- [Posts JSON API](https://arpitak.com/api/posts): Structured JSON list and full content of blog posts
- [Reading List JSON API](https://arpitak.com/api/reading): Structured JSON reading list
- [Highlights JSON API](https://arpitak.com/api/highlights): Structured JSON photo highlights
- [Now Playing JSON API](https://arpitak.com/api/now): Live or most recent Spotify track

## Social Profiles

- [X / Twitter (@arpitarunkumaar)](https://x.com/arpitarunkumaar)
- [GitHub (@arpitarunkumaar)](https://github.com/arpitarunkumaar)
- [LinkedIn (Arpit Arun Kumaar)](https://www.linkedin.com/in/arpitarunkumaar/)
- [Substack (@arpitarunkumaar)](https://substack.com/@arpitarunkumaar/likes)
- [Spotify](https://open.spotify.com/user/zp6ykurq8u1yy6qitl03dg5uc)

## Content Negotiation

Every content page on this domain supports `Accept: text/markdown` content negotiation (returning `Content-Type: text/markdown; charset=utf-8` and `Vary: Accept`) as well as direct `.md` URL suffixes (e.g. `/writing.md`, `/writing/ibert-fpga-impl.md`, `/reading.md`, `/highlights.md`).