Rushil Jaiswal

I build AI systems that have to be right about people.

Claremont McKenna ’29 · CS at Harvey Mudd · Claremont, CA

See the work GitHub Email

01  About

I patented an athletic recovery product at seventeen and shipped 600 units before I finished high school. That taught me the only lesson that transfers: the thing is not real until someone who owes you nothing uses it.

Right now I’m building Telos — AI candidate sourcing for IT staffing agencies. Most matching tools rank résumés by keyword overlap with a job description, which is a search problem pretending to be a hiring problem. Telos scores candidates against placement history: who actually got hired, for what, and what the recruiter said about them. My dad’s firm Flexton is the first design partner, which means my feedback loop is a recruiter telling me the ranking is wrong and why.

Before that: Vibecheck, which scans a repo with real static analysis and opens a PR that fixes what it finds; Strata, field guidance for plumbing contractors that has to keep working in a crawl space with no signal; and pipecat-firewall on PyPI, a drop-in guard that blocks prompt injection before a voice agent’s model ever sees the turn.

The through-line is systems that have to be right about people — matching them to work, remembering their lives, refusing to be talked into betraying them. In all three the interesting engineering is the failure mode, not the demo.

That is also why every project here ships with what is wrong with it. A portfolio of things that only worked is not evidence of judgement. Most of these pages end with the part I could not solve, and seven of them are dead.

Outside the terminal: competitive breaker, qualified for nationals. Guitar on Monday nights at Pomona. Currently training for a half marathon, badly.

02  Experience

Founder — Telos Current Jun 2026 — Building AI candidate sourcing for IT staffing agencies, solo — architecture, backend, frontend, and the client feedback loop. Live JobDiva sync, outcome-aware ranking, 23 merged releases, in the hands of a real firm.
VP of Partnerships — Claremont VC Current 2026 — Student venture group across the Claremont Colleges. I run partnerships — building the relationships with funds and founders that the rest of the club’s work depends on.
Founder — Wrapit Recovery Shipped High school Designed, patented and sold an athletic recovery product — 600+ units shipped. Manufacturing, fulfilment and customer support included, which is where most of the actual learning was.

03  Selected work

Telos Building 2026 — A résumé that echoes the job description is a search result. A person who gets placed is the answer. AI candidate sourcing that ranks on placement outcomes, not keyword overlap. Live JobDiva sync, 23 shipped releases, one real staffing firm using it. FastAPI · Pinecone · Supabase · GPT-4o · JobDiva API Goodhart Research 2026 If the grader can be gamed, training harder just produces a better cheater. RL verifiers are the bottleneck: train against a gameable grader and you get a reward hacker. A red-team swarm finds the holes, a green team seals them, and you get a before/after robustness number against held-out ground truth. Python · Anthropic SDK · FastAPI + WebSocket Envoy Hackathon 2026 An agent worth delegating to is one that holds your details under pressure. A voice agent you hand a real errand to: it calls the restaurant, negotiates with a live host and books the table. Because it is holding your details while it does, it also writes its own guardrail after a breach — 0/5 → 5/5, independently judged. Pipecat · NVIDIA Nemotron · Twilio · Cekura Hindsight Hackathon 2026 Memory teaches an AI your facts. This teaches it how you tend to be wrong. Memory teaches an AI your facts; Hindsight teaches it your patterns of being wrong. Two days later the concept shipped into GBrain as the named “Hindsight calibration wave” — the name is still in its schema migrations. TypeScript · Express · GBrain skills · Chrome extension Yaad Hackathon 2026 A fluent wrong answer about someone’s grandson is worse than admitting you don’t know. A bilingual voice companion for dementia patients, built on a memory graph that refuses to invent a life. Every answer is grounded in retrieved memory or it says it doesn’t know. LiveKit · Pipecat · MiniMax · Moss · FastAPI Vibecheck Shipped 2026 Real scanners find the bug. The model only translates it and writes the fix. Paste a repo, get a security report a non-engineer can act on — then one click opens the PR that fixes it. Real scanners, not an LLM guessing at vulnerabilities. Next.js · FastAPI · Semgrep · Gitleaks · OSV · Claude

All 23 projects, including the dead ones →

04  Stack

Languages

Python, TypeScript, SQL. Python for anything with data or a model in it; TypeScript when a person has to look at it.

Backend

FastAPI for services, Next.js when the front end and API belong together, Supabase/Postgres for state.

AI

Claude and GPT-4o, Pinecone for retrieval, Pipecat + LiveKit for voice. Deterministic code wherever a model is not actually required.

Quant

pandas, statsmodels, scikit-learn. Enough econometrics to know when a p-value is load-bearing and when it is decoration.

Infra

Vercel, Railway, Docker, GitHub Actions. Boring on purpose — deployment should not be the interesting part.

Reach for first

The smallest thing that can be put in front of a real user this week. Everything I have learned came from that, not from the architecture diagram.

05  Interests

  • AI agents under adversarial pressure
  • labor markets
  • calibration & forecasting
  • prediction markets
  • breaking
  • guitar
  • running
  • climbing

~/R