Llama Lounge Snowflake Agentic Hackathon: BarackOllama, CrewAI, and the Golden Llama

Mar 1, 2026

Following the deep-research marathons of late 2025 like the AWS & NVIDIA Hackathon and WeaveHacks 2, late February 2026 brought an exciting new challenge: the Llama Lounge Agentic Hackathon (February 27–28, 2026). I have always had a special affinity for two-day weekend hackathons—they pack an intense sprint of creativity, camaraderie, and problem-solving into a focused 48 hours.

Organized by Cerebral Valley in partnership with Jeremiah Owyang’s Llama Lounge, the hackathon was hosted at Snowflake’s newly opened Silicon Valley AI Hub (SVAI Hub) on Constitution Drive in Menlo Park. Snowflake opened its doors to the Bay Area’s top AI builders to celebrate expanding its AI footprint, and the venue was electric.

Being at Snowflake’s SVAI Hub was also a wonderful personal reunion. Former Google Developer Group (GDG) Fresno lead James Cha, whom I inherited the GDG Fresno organizer role from back in 2016, is now at Snowflake as Lead DevRel! It was fantastic to catch up with him right at the new hub. In fact, meeting James sparked an exciting plan: the next edition of our GDG Fresno tech tour to Silicon Valley may very well feature a stop at the Snowflake SVAI Hub!

Industry heavyweights and judges filled the floor, including Joao Moura (Founder & CEO of CrewAI), Palash Kala (Composio), Craig DeWitt & Ankit Agrawal (Skyfire), Umesh Unnikrishnan (Snowflake), and investors from AI Fund, Blitzscaling Ventures, and Breakthrough Ventures.


The Mission: Reimagining the Job Application Workflow

For this hackathon, I teamed up with three ambitious students eager to tackle a problem close to their hearts: the painful, high-friction job search process. My teammates were Ankita Sethi, Hima (Himavanth Karpurapu), and Shri (Shreyas Yadav).

In today’s hyper-competitive tech hiring landscape, blasting generic resumes into applicant tracking systems (ATS) is an exercise in futility. For every single role, candidates need to meticulously customize their materials:

  1. Surfacing Relevance: Reordering and rewriting bullet points so domain-relevant skills and past experiences appear front and center.
  2. Keyword Alignment: Intelligently aligning technical terminology with the nuances of the specific job posting without sounding artificial or stuffing buzzwords.
  3. Beyond the Static PDF: Modern developers have rich, living digital footprints—most notably their GitHub repositories. Yet candidates rarely, if ever, tailor their GitHub profiles or project showcases for a specific employer or role.

We asked: What if an autonomous multi-agent system could ingest a candidate’s background, analyze any live job posting, and autonomously tailor their entire application package—from a print-ready, keyword-optimized PDF resume to a customized GitHub README and a dedicated, role-aligned web portfolio?

Thus, BarackOllama was born. The name was concocted purely for fun—a playful pun blending Barack Obama and Ollama—and the organizers and judges definitely loved the humor!


Whiteboard Sketch to Multi-Agent Architecture

As Friday night gave way to early Saturday morning, we huddled around the whiteboard at Snowflake HQ to map out the complete multi-agent data flow:

(Late-night architecture brainstorm on the whiteboard at Snowflake’s Silicon Valley AI Hub.)

Translating this sketch into a production-grade multi-agent architecture required orchestrating several moving parts across data ingestion, reasoning, and deployment:

(Complete architectural flow of BarackOllama: scraping via Bright Data, multi-agent reasoning with CrewAI, tool orchestration via Composio, and artifact generation via WeasyPrint.)


Technical Highlights & The Stack

1. Data Ingestion: Solving the LinkedIn Wall with Bright Data

One of the earliest roadblocks in building career agents is data retrieval. Scraping LinkedIn directly—whether job listings or candidate profiles—is notoriously difficult due to aggressive bot detection, dynamic single-page hydration, and rate limiting.

Rather than building brittle custom headless scrapers that break halfway through a demo, we integrated Bright Data’s Web Scraping API. Bright Data provides dedicated snapshot datasets for LinkedIn job posts and member profiles:

  • We triggered asynchronous scraping snapshots for the target job URL and candidate profile.
  • Polled the snapshot API until data was ready.
  • Extracted clean, structured JSON containing the job title, formatted job description, seniority level, compensation range, and the candidate’s verified skills and work history.

2. Agentic Orchestration: Why We Chose CrewAI

With sponsors showcasing their latest agent platforms, we faced an architectural choice: use Snowflake’s integrated agent ecosystem (Snowflake Cortex) or CrewAI.

We opted for CrewAI, especially with CrewAI’s founder Joao Moura present in the room. CrewAI’s structured Flow and Crew abstractions allowed us to design a clean, modular multi-agent pipeline (ResumeFlow) with clear task handoffs and Pydantic-enforced schemas:

  • Research Crew (research_crew.py):
    • Keyword Extraction: Analyzes the scraped job description to extract high-priority technical skills, core qualifications, and domain competencies.
    • Portfolio Relevance Ranking: Compares the candidate’s public GitHub repositories against the extracted job requirements, ranking projects by relevance score and highlighting matching tech stacks.
  • Content Creation Crew (content_crew.py):
    • Crafts a modern, targeted GitHub profile README.md that elevates the candidate’s top-ranked projects for that specific role.
    • Generates a tailored single-page HTML/CSS portfolio website presenting the candidate’s most relevant work.
  • Resume Writing Crew (resume_crew.py):
    • An Executive Technical Resume Writer agent synthesizes all available context: the target job listing, LinkedIn history, ranked GitHub projects, and interactive links to the newly generated portfolio.
    • Generates a complete, self-contained HTML resume with embedded CSS optimized for print layouts using CSS @page media queries (A4/Letter dimensions, clean typography, clear section hierarchy).

3. Composio Tooling & GitHub Automation

Hackathon sponsor Composio provides over 800 ready-to-use APIs and MCP tools for AI agents. We leveraged Composio’s native CrewAI provider (composio_crewai) to seamlessly authenticate with GitHub.

Using Composio, our deployment crew could automatically:

  • Inspect the candidate’s public repositories and commit histories.
  • Authenticate via OAuth.
  • Programmatically provision a dedicated GitHub repository for the custom portfolio and commit the generated HTML/CSS artifacts and README directly to GitHub!

4. HTML to Print-Ready PDF via WeasyPrint

Generating resumes as raw text or standard markdown often yields clunky formatting when converted to documents. By having our Resume Writer agent output semantic HTML with print-ready CSS rules, we used WeasyPrint to render the dynamic HTML directly into a crisp, pixel-perfect, ready-to-download PDF resume.

5. Dual Interfaces: CLI & Web GUI

To ensure accessibility for both developers and non-technical job seekers, we built two working interfaces:

  • BarackOllama GUI: A full web interface and FastAPI server where users can input URLs, select options, track live agent progress, and preview their HTML/PDF artifacts.
  • BarackOllama CLI: A streamlined, developer-centric terminal tool experimenting with Google’s Agent Development Kit (ADK) and Gemini, executing the multi-agent pipeline right from the shell.

Project Submission & Demo Video

As summarized in our official project submission:

“Our project is a personalized, adaptive resume generation agent that automates the job application workflow. It uses Bright Data to scrape LinkedIn and GitHub data, leverages CrewAI for multi-agent reasoning, and integrates external tools through Composio to analyze job descriptions and perform intelligent, personalized keyword alignment with the user’s profile. The system then generates a dynamic HTML resume and converts it into a ready-to-download PDF, transforming structured data and agent-based reasoning into a real-world, customized application artifact.”

Here is our submission demo video walk-through, recorded by teammate Ankita Sethi:

This browser does not support the YouTube video player. Watch on YouTube

Demo walkthrough of BarackOllama automating role-tailored resume generation and GitHub portfolio customization.


The Golden Llama & The Sweet Outcome

The ultimate trophy of the hackathon was the legendary Golden Llama:

(The coveted Golden Llama trophy on display at Snowflake’s Silicon Valley AI Hub.)

While our team did not take home the Golden Llama or make the podium, the weekend yielded something even more valuable: admission into the highly coveted Llama Lounge event series.

For a long time, I had wanted to attend Llama Lounge events. Organized by Jeremiah Owyang, Llama Lounge is one of the most exclusive, high-signal event series in Silicon Valley, bringing together frontier AI startup founders, researchers, and pioneers under one roof. Competing at the Snowflake hackathon earned me access and an invitation into this community.

Since this hackathon, I have had the privilege to attend several landmark Llama Lounge gatherings—including unforgettable evenings at Google DeepMind and Stanford University. But those are stories for future blog posts!


Comments loading...