Ashleyn Castelino

Work & experience

2024 →

Applied AI Engineer & Lecturer UIUC Post-training an internal open-source GPT — SFT, RL, distillation — plus the Azure pipelines it ships on. Also teaching a course on becoming a quantitative analyst, combining AI, ML and finance.

→ architecture → course

+ what I built− less

— Our own Gies Chat model — post-training Qwen 3.5 into a tutor for business courses, so the model teaches the material the way the college does rather than answering like a generic assistant. Funded by a $5,000 research grant from Thinking Machines Lab. → model — Magelli Force — an operating system for the Magelli office: one place for staff to find, deliver and maintain client relationships. Search runs on a rank-fusion algorithm I built, retrieving across 200+ slides for users in production. → live → demo — Gies Chat — an open-source ChatGPT for Gies College of Business students. → live → demo → agent-builder deck — Azure migration — moved the college’s AI apps onto Azure and built the pipelines they ship on: container registries, model endpoints and staged deploys.

+ the rest− fewer

— Magelli Scout, inside Force — finds sales leads for the experiential learning office, scored against the Magelli rubric, with email automation and analytics on top. — Slate — project intelligence across a project’s whole life: developers keep working in GitHub, while managers and stakeholders get readable updates on progress, blockers, risks and decisions. — Atlas — a visual map of the academic portfolio, using agentic graph reasoning to connect departments and the courses they teach. — Venture Voice — an AI venture simulation that puts students through customer discovery before they talk to a real customer, over video and voice. → live — Viva — an AI-native assessment platform where students show what they’ve learned through video, presentations and oral exams instead of another written test. → live — ProjectClaw — an internal project assistant that keeps track of work across departments and inside teams.

may 25 →

Agentic AI Intern North Oil Company · Doha, QA LangGraph framework automating oil well-test analysis; RAG retrieval tuned for hallucination-free engineering advice.

+ what I built− less

— A multi-agent system in LangGraph that reads well-test data — bottom-hole pressure, oil and water production, water-to-oil cut — and recommends which production zones to keep open or shut in. — A planner splits the work after the pressure analysis, then specialist agents study oil trends, water cut and efficiency in parallel before a summariser reconciles them into one zone-by-zone call. — Tiered memory: runtime state in-process, semantic recall in a vector store, past evaluations and agent messages in a document store. — An evaluation harness scoring agent recommendations against engineer-labelled ground truth, so changes to the graph could be measured rather than argued about.

jun–aug 25

Graduate Research & Teaching Assistant UIUC · with Prof. Vishal Sachdev Researching and building tools that help students actually learn with AI, rather than hand the work to it — and TA’ing the course where they write the code.

+ what I built & taught− less

— TA for BADM 350, business technology — debugging students’ code with them, getting them building agents in Codex, and teaching the software concepts underneath it all. — TA’d two further courses — BADM 337, supply chain, and BUS 300, business consulting — managing their client projects: student teams working with real clients on real business problems, from scoping through delivery. — Researched drip learning — whether course material lands better delivered in small daily doses than in one sitting. — Built BADM Bot on the OpenClaw architecture to test it: a daily bot that drips one piece of the course to students and tracks what sticks, reaching them on WhatsApp instead of the LMS. → writeup — Built MindForum — an AI brainstorming tool that keeps a human in the loop rather than handing back a finished answer. → live → code

aug 24–may 26

Ventures

2025 →

Founder & Applied AI Developer Healthcare & edtech startups · Champaign & Chicago, IL Software development and customer discovery, engagement by engagement.

+ what I built− less

— Building agentic solutions for a healthtech startup — AI products that carry real clinical and operational work rather than sit beside it. — Founded an agentic CRM for customer relationships and workflows, plus an AI deployment platform for server infrastructure and production systems. — Built a healthcare AI eval pipeline measuring medical-document accuracy, source grounding and hallucination risk for a Chicago health startup. — Shipped an LLM-powered language-learning app for an edtech startup: personalized learning and user interaction.

2025–26

Founder & President Agentic AI @ UIUC · 500+ members · student organization Built UIUC’s first agentic-AI club and led it to 500+ members — the largest AI club on campus: lectures, hackathons, 30+ open-source projects.

→ club site → projects & slides → open-source projects

+ what I led− less

— Landed the club’s partnership with Optiver, its first industry partner. — Led the lecture series alongside an exec — technical AI, agentic systems, Claude Code and the OpenClaw workshops. — Built and led a technical exec team of five. — Owned the club’s project portfolio: hosting and maintaining open-source work built internally and with outside collaborators.

aug 25–may 26

Founder & President Macro Investing Association · UIUC · student organization Founded UIUC’s macro investing club: weekly market lectures, an education curriculum across every asset class, and a portfolio simulation competition.

→ blog → lecture slides (pdf)

+ what I led− less

— Ran an exec of five, six portfolio managers and 30 analysts, meeting weekly for market lectures. — Built and taught the club’s education modules across macro investing — equities, bonds, FX, commodities and crypto. — Taught technical analysis alongside the macro curriculum. — Built the club’s portfolio simulation competition, where members run their own book against the market.

jan–dec 25

Selected projects

click any card

Text2SQL Agent 2nd — Berkeley AgentX Catches phantom tables pre-execution. 2nd place, hardest track at Berkeley’s AgentX–AgentBeats.

→ deep dive → GitHub → press coverage

python / sqlglot / duckdb / sqlalchemy / flask / docker

ARC Capital personal project A macro hedge fund that runs 24/7: agents trade off fundamental and alternative data — Fed speeches, X and prediction markets — instead of waiting for a desk to open.

→ demo → deck → code

python / langgraph / typescript / react / docker

Resrv client project · solo · under 3 weeks Built and delivered to the stakeholders on my own, start to finish. Autonomous lab booking with human-in-the-loop overrides; 100+ users on demo day.

→ deep dive → short writeup → deck → code

python / fastapi / discord.py / typescript / sqlite

UrbanOps open source · with the club Agentic operations for city-scale workflows.

→ short writeup → GitHub

typescript / python / react / docker

+ show more− show less

Aegis Data Intelligence open source · MIT Agents that catch schema drift, diagnose root cause and propose the fix.

→ short writeup → GitHub

python / fastapi / langchain / typescript / react / docker

Slicky personal project Screenshot anything on your Mac, get an instant explanation — and a knowledge library that builds itself.

→ demo → code

typescript / rust / tauri / react / python

FocusTabs personal project Chrome extension that reads your open tabs against what you’re working on, suggests what to close, and archives everything so nothing is lost.

→ Chrome Web Store → GitHub

javascript / typescript / chrome mv3 / llm api

NexusAI Keywords AI (YC W24) hackathon · 24 hours Company knowledge scattered across GitHub, Jira and Slack, pulled into one queryable system. Graph traversal plus embeddings — more grounded than either alone — answered by voice or chat, with a team-analytics view on top.

→ short writeup → code

python / typescript / graph + embeddings

CogniMap VisaVerse AI hackathon · team of 2 Course platforms teach everyone the same way. CogniMap reads what you already know from your background and career, then generates the modules for the gap that’s left.

→ Devpost → code

typescript / gemini / google ai studio

Super Ralph personal project An autonomous agentic loop for Claude Code: it decomposes a request into tasks, writes strict tests first, implements with fresh sub-agents, debugs itself cold when stuck, and logs what it learned from every run. → code claude code plugin / agents / shell

SonicScout personal project An AI-native venture analyst: onboard a fund’s thesis once, hand it a startup URL, and five specialist agents return an investment memo in about 90 seconds — pass, watch, take the meeting, or invest.

→ short writeup → code

typescript / agents / web3

MedAgent IEEE SIEDS 2026 · peer-reviewed Health search with citations, confidence scores, and stated uncertainty.

→ slides (pdf) → read the paper → code

python / pgvector / postgres / openai / networkx

Press & mentions

2025 → 2026

Runner-up honors in a major agentic AI hackathon Gies College of Business · news Coverage of Text2SQL taking second in the hardest track at Berkeley’s AgentX–AgentBeats. → read the article 2026.04

AI for Impact Challenge at Gies College of Business

AI for Impact Challenge (hackathon) Gies College of Business · news The 24-hour agentic-AI challenge at UIUC, opened by Decasonic’s Paul Hsu — the college’s writeup of the systems teams built for its own operations.

→ read the article → see the post

2026.08

Invited lecture on global financial markets, SJEC India

Invited lecture on global financial markets St Joseph Engineering College · MBA The institution’s writeup of the lecture on global macro, emerging markets and tradable instruments. → read the writeup 2026.01

+ other mentions− less

Second place at the Claude Builder Club hackathon LinkedIn · post by Keshav Dalmia A teammate’s writeup of the MCP server we built in three hours at UIUC — long-term memory, a knowledge graph over team conversations and an analytics dashboard on top. → see the post 2025

Writing

linkedin · substack · x

tech

Why you need a personal benchmark for your AI 2026.10

Automating a client’s Salesforce workflow 2026.09

GPT: a genius at code, with no sense of taste 2026.09

Applied AI is the theme for the rest of the decade 2026.09

Money for machines — agentic payments 2026.09

+ 30 more in tech− fewer

NVIDIA’s open agent safety platform 2026.09

Meta’s Muse and the personal agent hype 2026.09

We’re confusing capability with intelligence 2026.09

RLCD: teaching agents to know when they’re right 2026.09

Frontier labs vs open-source labs 2026.09

Inside the Astra hackathon in NYC 2026.09

Applied AI doesn’t work unless you think in workflows 2026.09

Putting Astra’s 3D video to the test: UIUC in 3D 2026.09

Abu Dhabi enters the AI race: K2 Horizon and sovereign AI 2026.09

My most-used AI coding skill: mini-brainstorm 2026.09

The next generation of models is here 2026.09

How GLM-5.3 was post-trained 2026.08

GLM-5.3 Flash, an open MoE for coding agents 2026.08

Sales agents that learn from feedback 2026.08

Meta’s open-source return, post-trained for agents 2026.08

The future of enterprise AI: own the model, own the harness 2026.08

Why agentic AI needs a different learning loop than coding 2026.08

Before continual learning, LLMs need to learn from their mistakes 2026.07

How Kimi K-3 was post-trained 2026.07

Why small businesses don’t buy AI (for now) 2026.07

Claude Opus 5 and cost per intelligence 2026.07

On Meta’s Muse Spark 2026.07

The cost-per-intelligence frontier 2026.07

Loop engineering 2026.07

Fine-tuning a 7B model for agentic deep research 2026.07

Notes on coding agents 2026.07

Fewer tools, better agents 2026.06

Agentic coding, in practice 2026.06

Post-training agents with agentic RL 2026.06

Reading DeepSeek V4 2026.01

finance substack →

This week in markets and AI — rate hikes return, Meta’s agent play, and Anthropic’s IPO 2026.09

This week in AI and markets — inflation, $100 oil, and the model race 2026.09

The jobs print that put a hike back on the table 2026.09

Healthcare is back? 2026.09

Fed interest rates, and what they mean for AI models 2026.08

+ 11 more in finance− fewer

Market wars, from Japan to small caps 2026.08

Oil past $100, and the technicals behind it 2026.07

Hyperscaler debt 2026.07

Lower-end bonds are a story 2026.06

Market dip, really? 2026.06

MIA Newsletter #9 — Brazilian real, the shutdown deal, crypto & gold 2025.11

MIA Newsletter #8 — equity correction and a volatility spike 2025.11

MIA Newsletter #7 — FOMC, the trade deal, and October’s rebound 2025.11

MIA Newsletter #6 — CPI, India–US trade, Argentina’s midterms 2025.10

MIA Newsletter #5 — Fed cut, yields at one-year lows, gold over the S&P 2025.10

MIA Newsletter #4 — 100% China tariffs, shutdown stalemate, record gold 2025.10

Online books

interactive textbooks

Global Macro, From Scratch Macro investing · free & open A free, interactive textbook on global macro investing, starting from what an investment actually is. A start-here primer, then five parts: how the economy works, the policy makers, the asset classes, the global system, and doing macro yourself. → read the book

2026

AI, ML, Markets Textbook UIUC · MSBA An interactive textbook for the Quantitative Analyst Workshop Course — nine chapters in four parts, set at a fictional equity research firm adopting agents one step at a time. Every chapter has Python that runs in the page, a visual explainer, a checkpoint quiz and a Colab notebook against the real API. → read the book

2026

Talks & workshops

uiuc · gies · mia · agentic ai

tech

Quantitative Analyst Workshop Course UIUC · MSBA · weekly A workshop series on becoming a quantitative analyst — AI, ML and finance taught together. Every session, recorded and with its deck, sits on one page.

→ every session → course site

2026–

Prompt to harness engineering — lecture YouTube · recorded lecture Teaching the move from prompt engineering to harness engineering — when the unit of work stops being the prompt and becomes the system the model runs inside.

→ watch the lecture → slides (pdf)

2026

AI for Impact Challenge (hackathon) — host & organizer Agentic AI @ UIUC · Founder & President 24 hours, 5 tracks, real business problems — opened by Decasonic’s Paul Hsu. — Ran 5 workshops in the lead-up, from AI foundations through building multi-agent systems. — Built HackClaw, an agent that ran the event itself — logistics, participant queries and the technical questions teams hit mid-build.

→ press coverage → event recap → opening slides (pdf)

2026

+ read more− show less

OpenClaw × Claude Code workshop Agentic AI @ UIUC · Founder & President The largest OpenClaw meetup at UIUC. With BuildIllinois: 70+ students each shipped a NanoClaw — a lightweight OpenClaw of their own — deployed live, with $1,000 for the best build.

→ event recap → slides (pdf)

2026

Agentic AI workshop — UIUC Invited by UIUC · education administration Invited by UIUC to brief its education administrators on agentic AI — what the systems actually do, and what they change for how the school works.

→ watch the workshop → slides (pdf)

2026

Memory management — lecture Agentic AI @ UIUC · Founder & President How agents remember: what to keep, what to summarise and what to throw away — and why most agent failures start here. → slides (pdf)

2026

Agentic coding workshop Invited talk · entrepreneurship course, UIUC Invited into an entrepreneurship course to explain AI and agentic coding to founders: where coding agents stop being autocomplete, practical workflows, and live demos of Cognitive Swarm, HackClaw and BADM Bot. With Keshav Dalmia.

→ recap on LinkedIn → slides (pdf)

2026

The evolution of AI agents & RL Invited talk · UIUC Invited by UIUC to trace agent architectures end to end — from prompting through reinforcement learning — in one workshop. → slides (pdf)

2026

GenAI architectures — lecture Agentic AI @ UIUC · Founder & President Context engineering, end to end: RAG, MCP, single- vs multi-agent, and a teardown of how Copilot assembles context. Deck and code shared with the club. → lecture recap

2025

finance

Weekly market update ARC Research · Substack · weekly A weekly read on global markets and AI — central banks, rates, oil and the model race — written up on Substack and presented from a deck. Every update, with its deck, sits on one page.

→ every update → subscribe

2026–

Global financial markets: an overview Invited lecture · St Joseph Engineering College, India Invited by the MBA department: the global macro environment, emerging markets and tradable instruments, and why bond markets matter more than the room expected.

→ SJEC writeup → slides (pdf)

2026

FX & crypto — lecture Macro Investing Association · Founder & President Carry trades, historical FX blowups, and how crypto actually moves with macro.

→ recap on LinkedIn → all MIA lecture slides (pdf)

2025

+ read more− show less

Market discussion — lecture Macro Investing Association · Founder & President The 50-year mortgage proposal, December rate-cut odds, and EM allocation — plus portfolio program updates.

→ recap on LinkedIn → all MIA lecture slides (pdf)

2025

Commodities — lecture Macro Investing Association · Founder & President Hard and soft commodities under macro pressure, read through the 1973 oil crisis and the Hunt brothers.

→ recap on LinkedIn → all MIA lecture slides (pdf)

2025

Open market discussion — lecture Macro Investing Association · Founder & President Ran it as an open floor: US–China trade, oil in the $60–65 band, and what the pods got wrong that week.

→ recap on LinkedIn → all MIA lecture slides (pdf)

2025