notebook no. 3 · toronto, on · utc−5
Rajan Chavada
I build tools that engineers actually keep using.
Right now I'm on the backend at Bree (YC). Before that I was moving models onto GPUs at RBC Borealis AI, and before that I built a research agent at RBC that ended up patent-pending. Nights and weekends go to Neurovn, CacheLane and Rosetta. I study CS at Western, class of '27.
Things I've built
(and still maintain)- part no. RC-001 · status: live alphasheet 1/4
Neurovn
Solo buildKnow what your agent costs before it runs.
Drop agents, tools and conditions on a canvas and Neurovn estimates token cost and latency for the whole graph, loops included, before a single API call happens. It covers 3,500+ models across 700+ providers, and a Python SDK traces real runs so you can check the estimate against reality.
neurovn demo, no audio, real footage - proof
- Inbound interest from Google Cloud engineers.
- made of
- React · TypeScript · Node.js · Python SDK · graph algorithms
- links
- live app ↗github ↗pypi ↗
↳ started because nobody at work could tell me what our agents cost
- part no. RC-002 · status: on npmsheet 2/4
CacheLane
ShippingStop coding agents from burning your API budget.
A tiered, fail-open caching layer for Claude Code sessions. It sits between the agent and the API, prunes and caches what it safely can, and gets out of the way when it can't. Co-designed with Aditya Tripuraneni; the team uses it daily.
cachelane demo, no audio, real footage $ npm i -g cachelane && cachelane install - part no. RC-003 · status: on npmsheet 3/4
Rosetta
AdoptedWrite your agent rules once. Use them in 9 IDEs.
A CLI that translates agentic-coding configs between Cursor, Claude Code, Windsurf and six others from one master spec. I pitched it at Borealis AI and other engineers on the team picked it up.
$ npx rosettablueprint init - part no. RC-004 · status: internalsheet 4/4
Agentic research at RBC
Patent pendingSEC filings and live news, read for you before the client call.
A multi-agent research system for hedge-fund coverage at RBC Amplify. It pulls filings and news, reconciles them, and hands traders a brief. It's internal, so there's no demo, but the patent application covers how it's orchestrated.
- proof
- Trader prep time went from 6 hours to 2.
- made of
- LangChain · FastAPI · React · Docker · OpenShift
- links
- the patent story ↗
also on the bench · hackathons & older builds
- Badgerésumés → vectors → a 3D map of who you should talk to at a career fairUofT Hacks winner
- PhysioPointARKit app that measures joint angles like a physiotherapist wouldSwift Student Challenge
- NVIDIA alert triagemulti-agent on-call engineer for GPU clusters on Nemotron-70B
- Plycelocal restaurant discovery for iOS
- Chill Billfinancial wellness app for Gen Z
- ASL translatorCNN sign-language translator, Western AI team
Work log
six internships, zero coffee runsSep 2026 – now
Bree/ Y Combinator
Software Engineer Intern
One of two backend engineers on a small founding team. I own the weekly release train: CI/CD, canaries on Lambda and API Gateway, and whatever shows up in Datadog at 11pm.
★ p99 API latency down 35%
- Staged rollouts route 40%+ of live traffic through canaries, so 100k+ users never see a bad deploy.
- Tracked down latency bottlenecks across the TypeScript and Python backend without breaking the financial math.
- Deploys went from ~5 minutes to under 1.
Jan – Aug 2026
RBC Borealis AI/ RBC's AI research lab
ML Software Engineer Intern
Sat between researchers and GPUs. Moved credit-capacity model retraining onto NVIDIA Triton and built the dev environment every researcher on the team now starts from.
★ 12GB model → 170MB (70x smaller)
- Inference time cut 44% after the Triton migration, working directly with researchers on architecture changes.
- Built the devcontainer system from scratch: multi-hour manual setup became one-click GPU provisioning.
- Distributed training pipelines on PyTorch + RUN:AI so compute gets allocated instead of fought over.
- Pitched Rosetta to the lab; other engineers on the team now use it.
Sep – Dec 2025
Intact Insurance
Site Reliability Engineering Intern
Wrote Terraform so app teams could stop filing tickets for infrastructure and just provision it themselves, compliant by default.
★ provisioning time down 65%
- Reusable AWS modules for networking, permissions and access control.
- With a senior security engineer, turned a manual compliance review into a pipeline that clusters error patterns across the container fleet, reaching 95% compliance.
May – Aug 2025
RBC Amplify/ Hedge fund research
AI Engineer Intern
Built an agentic research system that reads SEC filings and live news so traders don't have to before a client meeting. It's now patent-pending.
★ trader prep: 6 hours → 2
- LangChain + FastAPI backend, React/TypeScript frontend, Dockerized on OpenShift.
- Most of the work was turning vague stakeholder asks into a spec people could actually sign off on.
- I wrote about filing the patent at 21 on Medium.
May – Aug 2024
RBC Capital Markets
Software Engineer Intern
Built the React component library traders now see every day, plus a ranking service that flags which news actually moves the market.
★ 30+ components, 100+ daily traders
- Reusable React/TypeScript components used across multiple business lines; page loads got faster along the way.
- Regression-based ranking service on Snowflake + ClickHouse, served via FastAPI to 3 trading desks, 95%+ accuracy on signal-to-movement for high-impact news days.
May – Aug 2023
RBC Global Equities
Software Engineer Intern
My first internship. Built a CVE triage platform so engineers could see which vulnerabilities actually mattered, instead of scrolling a spreadsheet.
★ triage time down 40%
- React + FastAPI, auto-ingesting MySQL and JIRA on a schedule.
- Adopted across 3 trading desks and 100+ Capital Markets engineering teams.
Fixes in other people's code
all merged- MERGEDcrawl4ai73k★deleted 1,900+ lines of dead code (an unreachable duplicate file and two orphaned functions)
- MERGEDhaystack26k★four test cases were silently sharing one pytest ID; gave them their own
- MERGEDopikBedrock pricing lookups broke on version-pinned model names; added a regex fallback
- MERGEDcamel-aiAnthropic rate-limit errors no longer crash the ChatAgent retry loop
- MERGEDlangfuseremoved retired feature-preview plumbing
toolbox
- write in
- Python, TypeScript, SQL, Java, Swift
- build with
- React, Next.js, FastAPI, Node, LangChain, PyTorch
- ship on
- AWS, GCP, Docker, Kubernetes, OpenShift, Terraform, GitHub Actions
- watch with
- Datadog, Grafana, Prometheus, NVIDIA Triton, RUN:AI
Writing
when I have something to say- Apr 2026140 Engineers, Not One Knew What Their Agentic Workflow Was Costing.Why I started building Neurovn, and what happened when I asked a lab of 140 engineers a simple question.medium ↗
- Apr 2026I Filed a Patent at 21 During My Internship.What nobody tells you about the process: the legal back-and-forth, the politics, and the writing that actually matters.medium ↗
- Dec 2025How I Use AI as a DeveloperThe context I give models, what I leave out, and where I stop letting them drive.medium ↗
- Mar 2025The Hidden Costs of Leaving Cash Sitting in Your BankAs a 20-year-old student living on campus, I have to juggle tuition, rent, and everyday expenses. But no matter how much I save, my money shouldn't be losing value.2 min