KaranKapur

Five projects

• Apr 2026 to present Open-source LLM inference infrastructure

  thaw

  Co-founder and cloud storage engineer. Rust and CUDA.

  • Cut LLM session cold boots on H100s about 386x, from about 340 s to a 0.88 s median fork. Receipt.
  • Own the cloud storage layer that persists and restores live LLM state: weights, KV cache, prefix-cache metadata and scheduler state.
  • Co-founded with N. Matteson and M. Yu.

  github.com/thaw-ai/thawthaw.sh

• Jul 2026 Agentic decision engine, solo build

  RelayIQ

  Decides whether a lead is worth enriching before you pay for it.

  • 3.3x lower cost per usable lead on a 495-case seeded benchmark: a multi-step agent routes each field to the best-fit enrichment provider and only spends a credit when it's worth it.
  • Raised precision from 0.682 to 0.773 across six strategies: calibrated confidence (Brier 0.164, ECE 0.091) holds low-confidence writes back from HubSpot and escalates them to a human in real time.

  github.com/karank2512/RelayIQ

• Sep 2026 Agentic workforce platform, solo build

  Foreman

  Describe a job in plain English and Foreman designs a worker for it. Open source, bring your own keys.

  • Lowered cost per run 30% vs an all-LLM pipeline on a 100-record paired benchmark: routed 5 of 7 steps to deterministic code, taking LLM calls per run from 11 to 6 (list prices).
  • Prevented repeated side effects in all 10 tested crash and race failure modes: a Postgres-backed run queue with leases, heartbeats, crash recovery, and idempotent resume restores a crashed run within 3 minutes.
  • Every tool call behind server-side permission checks and every external action behind human approval: an AES-256-GCM credential vault with key rotation, SSRF and DNS-rebinding guards, spend caps.

  github.com/karank2512/Foreman

• Oct 2026 Explainable buying-signal pipeline, solo build

  tell

  Who's hiring for the problem you solve?

  • Shrank entity-resolution review load 85% (625 to 91) as auto-link precision rose from 8 of 21 to 7 of 7 on 16,243 held-out SEC Form D filings: guarded name matches by issuer type, incorporation date, and fund status.
  • Monitored 907 companies and 10,585 open postings at 97.8% run success (321 runs) for $0/month: 6 collectors on a 4x/day GitHub Actions cron into Neon Postgres; sha256 event IDs make reruns insert 0 rows.

  github.com/karank2512/tell-aigithub.com/karank2512/tell-data

• Oct 2026 Agent telemetry transport verifier, solo build

  Delivery Doctor

  Your agent sent the event. Your dashboard never got it. This names the hop it died at.

  • Named the exact hop where agent telemetry dies in transit across 9 reproduced failure modes, from buffer drops and expired keys to 429s and queue overflows: replayed tagged canary events through the SDK and read each one back from intake.
  • Turned client-side diagnosis into a lookup, not a guess: a delivery ledger recording every event state change, lined up against the server's intake log. Zero runtime dependencies, one local command.

  github.com/karank2512/delivery-doctor

Background

Where I've worked

1. Aug 2026 to Present Data analyst Wisconsin Union. Madison, WI
   • Replaced the team's old scheduling software with an app I shipped in 2 days, used daily by 250+ employees: 4,400+ shifts, self-serve accounts, and work chat; built on TypeScript, MySQL, Docker, and AWS.
   • Forecasted weekly food demand and staffing from sales and shift data in Excel to guide orders and schedules.
2. Apr 2026 to Present Co-founder and cloud storage engineer thaw. Madison, WI
   • Cut LLM session cold boots on H100s about 386x, from about 340 s to a 0.88 s median fork. Co-founded the open-source inference startup (Rust, CUDA) with N. Matteson and M. Yu.
3. Jan 2026 to May 2026 Software engineer, senior capstone Eurofins Scientific. Madison, WI
   • Shipped a forecasting product lab technicians used daily: partnered with Eurofins as a client through UW-Madison, interviewed technicians and chemists across 4+ business units, and took it to production.
   • Turned 10K+ messy records into forecasts with confidence scores: built the ETL pipelines, PostgreSQL schema, FastAPI backend, and React front end; ran sprints as Scrum Master.
4. Jun 2025 to Aug 2025 Software engineer OpenAI x UW-Madison SAIL Program. Madison, WI
   • Reduced manual lookup time 40% and cloud cost 25% for 60+ users in UW-Madison IT: built a GPT-4 RAG platform serving 50+ concurrent queries at sub-2s latency.
5. Jan 2025 to May 2026 Computer support assistant UW-Madison, Department of Electrical and Computer Engineering. Madison, WI
   • Resolved 1,000+ time-sensitive issues at 98% satisfaction; monitoring automation lowered downtime 40%.
6. Jul 2024 to Sep 2024 Automation and compliance engineering intern HCLSoftware. Noida, India
   • Lifted onboarding conversion 30% via an analytics-driven redesign; 50+ tests shortened regression 60%.
7. Mar 2026 to Present VP, National Executive Council Kappa Eta Kappa
   • Converted a 50+ school TAM into 2 chartered and installed chapters: built a Python scraping pipeline that found universities without a professional engineering fraternity, sourced contacts, and ran the outreach.
8. May 2024 to Dec 2024 President, Delta chapter Kappa Eta Kappa, Madison
   • Led a 50+ member chapter with a $100K+ annual budget and an 8-member executive board, with events drawing up to 500 attendees. Liaison to the national council and alumni boards.

B.S. Computer Science and Data Science, Minor in Entrepreneurship. University of Wisconsin-Madison, May 2026. Coursework: Artificial Intelligence, Big Data Systems, Operating Systems, Algorithms.

Skills

• Languages and stack

  Python, TypeScript, SQL, Rust, C++, Next.js, React, FastAPI, PostgreSQL, Docker, AWS

• AI and agents

  LLM tool calling, agents, RAG, LLM-judge evals, calibration, Vercel AI SDK, MCP, vLLM, CUDA

• GTM and tools

  HubSpot APIs, webhooks, Clay and Snowflake (familiar), Claude Code daily, Cursor