~/vedant_
homeworkprojectsthoughtsabout

# work

what i ship.

i keep finding myself at very small companies right when they're deciding what to build. currently a stealth-stage ai startup; previously runanywhere (yc w26) and tegore-ai (yc x25). early-stage is where i'm sharpest — small teams, real users, weekly cadence.

the unstructured version of the same restless mind lives at /about; long-form project writeups at /projects.

↗ resume.pdf·github·linkedin·email

$ ./now● live

Stealth Startup

Software Engineer · Feb 2026 → present

San Francisco · forward-deployed at an enterprise client

Building an end-to-end multi-agent AI platform that autonomously processes insurance claims — voice calls, browser automation, and LLM reasoning, orchestrated.

  • LLM orchestration layer that loads claim state, plans next actions, and dispatches to specialized agents (voice / browser / sheets). Durable execution on Temporal with Redis-backed distributed locking and retry for multi-call campaigns.
  • Real-time voice pipeline on ElevenLabs Conversational AI + Telnyx SIP: live payer calls with claim context dynamically injected, structured outcomes parsed back from transcripts.
  • Vision-language browser agents on Playwright that drive insurance portals (Availity, Kern), fill multi-step forms, and programmatically clear MFA via the Microsoft Graph API.
  • Forward-deployed at an enterprise client — owning the relationship and integrating the pipeline into legacy order processing.
TypeScriptPythonTemporalRedisPostgresPlaywrightVLMsElevenLabsTelnyx SIPMicrosoft GraphDockerAWS

# previously

past full-time and internships, newest first.

  1. RunAnywhere (YC W26) · Software Engineer Intern

    Dec 2025 — Feb 2026

    San Francisco, CA

    Local-first inference SDK — bringing real models to the browser and the edge instead of the data center.

    • Extended the core inference SDK with multi-provider model routing, optimized tokenization, and request batching. LoRA-fine-tuned models down to a footprint that fits edge devices.
    • Built the on-device browser agent — VLM-driven web automation in a Chrome extension with Transformers.js, ONNX runtime over CDN, and a state-machine-first architecture that survives restricted pages and MV3 CSP.
    • Shipped Android Use Agent into the SDK Playground — VLM-on-Android wired into a real interaction loop, plus x86_64 ABI support so it runs on emulators, not just hardware.
    • Cross-platform Playground (Swift + Android + on-device browser) so partners and customers could touch the SDK without standing up infra.
    TypeScriptSwiftKotlinWebGPUTransformers.jsONNXVLMsPyTorchLoRA

  2. Tegore-AI (YC X25) · Software Engineer Intern

    Oct 2025 — Dec 2025

    San Francisco, CA

    AI tutoring that draws on a whiteboard mid-conversation. Voice and text in the same session, sharing state.

    • Designed an LLM tool-calling architecture that dynamically renders interactive React components mid-conversation — so the tutor can draw a diagram or quiz the student without leaving the chat.
    • FastAPI / TypeScript / Postgres backend with prompt-chaining, contextual retrieval, and concurrent workers. +27% tutor accuracy, −35% latency.
    PythonTypeScriptFastAPIPostgresOpenAIElevenLabsCartesia

  3. The Mind Company · Machine Learning Engineer Intern

    Sep 2024 — Dec 2024

    San Jose, CA

    Real-time brain-computer interface — EEG signals to motor-imagery control, sub-50ms.

    • CNN + Common Spatial Patterns classifier hit 92% on a 4-class motor imagery task. INT8 quantization let it run on a Raspberry Pi at sub-50ms inference.
    • Rebuilt the signal-processing pipeline with ICA artifact rejection and adaptive bandpass filtering — +12 dB SNR. Mixed-precision PyTorch training cut iteration time 3×.
    PythonPyTorchNumPyEEG / DSPRaspberry PiQuantization

# say hi

building something where this résumé would be useful?

i'm most useful around multi-agent systems, voice + browser automation, and infra for AI that has to actually run in production. write to me.

send me an email