Homen Shum

Joined Cheiron as its 26th team member following the company’s $8M seed round led by Menlo Ventures, bringing total funding to $13M. Cheiron is building an AI-native operating system for drug programs, connecting scientific, clinical, regulatory, and commercial evidence so biopharma teams can understand and execute complex workflows across the drug-development lifecycle. Working on the backend and applied-agent systems that power these workflows, including: - Building multi-step agent workflows for real biopharma customer use cases such as change requests and cross-region impact assessment. - Developing backend services, tool-calling systems, retrieval and evidence pipelines, and human-review workflows. - Grounding agent outputs in source documents, policies, and traceable evidence while keeping critical execution and approval boundaries deterministic. - Working directly with engineering, product, design, life sciences, business teams, and pilot customers to turn regulated workflows into production features. - Helping improve the reliability, evaluation, observability, and iteration speed of Cheiron’s applied-agent platform.

Started as a personal AI research project in December 2022 and incorporated as CafeCorner LLC in February 2025. - Shipped NodeBench AI, a Convex-backed entity-intelligence workspace for sourced research, reusable reports, agent chat, inbox workflows, evidence boundaries, evaluation, and a hosted public-research MCP. - Built NodeRoom, a collaborative human-agent workspace for spreadsheets, notebooks, decks, graphs, and other durable artifacts; proposals, version checks, and semantic rebase preserve human edits. - Distilled NodeAgent as the portable runtime and NodeKit as the application-composition, conformance, evaluation, and proof layer for the broader ecosystem. - Built NodeProof, NodeTrace, NodeMem, NodeRL, NodeSlide, NodeVoice, and supporting evaluation and product-development methods. - Published reusable engineering methods including agentic-ui-qa, Solo Founder Agent Builder, NodeAgentSpec, FeatureClipStudio, BetterPRHandoff, and NodeSEO.

Embedded at Meta Product Quality & Experience from November 2025 through March 2026, then continued with Tests Assured's TA Studio team. - Built 27 Python-backed agent tools spanning task intake, Android control, UI observation, authentication, state recovery, and verdict recording. - Returned structured receipts from tool adapters, linking each action to its observation, assertion, evidence, and verdict. - Built a multimodal Think -> Act -> Observe pipeline and bounded self-correction across 240 tasks, improving benchmark completion from 4/13 to 9/13 against a 17% recall / 39% precision baseline. - Applied TOON, a UI-hierarchy compression format that reduced device payloads 30-60% and long runs from 10-50M to 1-3M tokens. - Built TA Studio's device-emulation and QA MCP platform after the Meta engagement, reducing rerun cost 98% and reaching 100% precision, 80% recall, and 0.89 F1 on planted bugs.

Core engineer working on agentic manual-test-generation and self-adaptive UI/UX testing for Instagram, Facebook, and WhatsApp. - Designed agent architecture and evaluation across 214 bug-reproduction and 26 test-generation tasks. - Built mobile-control, state-observation, evidence-capture, and self-correction workflows using Python, MCP, LangGraph, Playwright, and ADB. - Converted multimodal Figma and product requirements into manual test cases through parallel visual and logic agents.

- Shipped MewAgent, an editor-native command surface with schema-validated Google GenAI and OpenAI tool calls, database mutations, and context-aware search inside a PKM notebook. - Built CI/CD regression gates for F1, latency, and tool-call accuracy using frozen golden datasets. - Reduced RAG embedding cost 80-90% through lazy indexing and batching; built an HTML parser that preserved nested web outlines during import.

- Supported prospect identification, account opening, KYC coordination, onboarding, and banking partnerships for healthcare and life-sciences founders. - Built LLMsuite, a Python/Streamlit agentic research workflow connecting PitchBook, CRM, company data, and 100,000+ documents while preserving human review. - Reduced manual entry 92%, sector classification from hours to 1-2 minutes, and prospecting from four hours to 15-30 minutes. - Compressed a 2,000-company money-movement analysis from three weeks to under 15 minutes. - Refined healthcare classifications using FDA pathway, payer-adoption, and commercialization criteria; presented startup-banking, cap-table, and QSBS material at founder programs, resulting in 32 follow-up signups.

- Designed and demonstrated an early agentic RAG workflow to the Palo Alto AI/ML team, supporting multi-entity comparison, dynamic tool use, structured extraction, and self-validation. - Translated banker and analyst workflows into requirements for AI/ML and intelligent-document-processing teams; built Python extraction, prospect-selection, and automatic color-coding tools. - Led cross-functional design of the Bulk List Extract capability in go/Companio, reducing manual extraction approximately 50% and projecting more than 100 staff hours saved monthly. - Led JPMorgan's Palo Alto AWS DeepRacer team to the #1 Silicon Valley regional rank and a global-finalist appearance; reduced model-training compute cost 15.9x through reward engineering and SageMaker optimization.

- Supported 72 credit transactions representing up to approximately $800M in aggregate exposure across technology, e-commerce, healthcare, and business services. - Built or updated financial models, credit packages, covenant analyses, annual reviews, and underwriting materials. - Evaluated historical and projected cash flow, downside cases, EBITDA adjustments, FCCR, leverage, liquidity, borrowing bases, covenant headroom, and startup runway. - Supported $10M-$30M facilities with startup-appropriate covenant and monitoring structures and contributed to multiple $100M+ transactions. - Mentored analysts with a reusable financial-modeling guide and presented risks and structures to senior stakeholders.

- Managed analysis across 15+ transactions during an intensive 3.5-month period and prepared due-diligence lists for underwriters and associates. - Consolidated banker, analyst, client, and credit-officer inputs into clear decision materials. - Supported client communication, commercial-card upsizes, preliminary 270 models, total-addressable-market analysis, and long-timeline healthcare and life-sciences transactions. ## Projects To Add or Refresh

Co-founded healthcare AI startup deployed on Azure [secured $500K Azure credit sponsorship]. Built 3 production prototypes: - GPT-4 SOAP note synthesis -- real-time clinical note-taking with HIPAA-aligned serverless architecture. - ICD-10/CPT code recommendation via vector semantic search over CCSR medical taxonomies -- shipped a RAG coding assistant that proposes diagnostic and procedure codes with precise citations. - Multilingual patient screening interface with PubMed research integration. Technical stack: Azure [Functions, Event Grid, Key Vault] ensuring full HIPAA compliance. Built a Streamlit ingestion pipeline with Docling and PII masking for secure data processing.

Build risk management skills while managing the daily fluctuations to less than 1% +/- of a $100k portfolio. December 1st, 2020, until February 28th, 2021

Generated 33% total profit within 20 market trading days by November 31st, 2020 Demonstrated proper utilization of risk management strategy, understanding of moving averages, and market correlations to maximize leverage using equity options

Researched companies with large market capitalization with analyst reviews and consumer reports from financial news resources and online product reviews to determine consumer enthusiasm and company’s growth potential Interpreted company performance with earnings growth delta filed on SEC EDGAR, modeled on Microsoft Excel

Collaborated with on-campus organizations to plan professional networking events and discuss marketing strategies Facilitated internal meetings, delivered recommendations to increase meeting efficiency and cut down meeting time Built an extensive network to organize and promote a quarterly mentorship program for new prospects

Planned, scripted, and facilitated a fraternal mentorship pairing event with a committee Booked rooms and requested fundings by outreaching the student body Managed sign-ups, flyers, tabling, and actively involved in every recruitment event

Invited a net worth $100,000 start-up organization to speak at the professional event held for the professionals at Delta Sigma Pi meeting.

Supervised members with up to $2000 from AS EOP Grant, Emergency Grant, CFF Grant, Young Entrepreneur Scholarship Provided online academic counseling with major selection and facilitated academic study groups Hosted professional development workshops such as resume and cover letter workshops

Trained a machine learning predictive model that can classify health status based on gender, eating habits, and physical condition using pandas dataframes and the sklearn ML classification algorithm Utilized data visualization tools such as Matplotlib and Seaborn pair plots to identify key factors used in training the model Collaborated with the front-end development team to host the algorithm on Google Cloud Platform and create a REST service using Flask. Developed a web and mobile-friendly website using React, Firebase, Axios, and Material-UI that allows users to track their health status records on their Google Account Finished in the top 5 teams out of 257 developers of all skill sets in a competition, and presented the website and model to guests including the Amazon Alexa team, Appfolio, and Invoca.

Used excel sheet to sort consumer transaction and demographic dataset Created PowerPoint presentation to outline a systematic approach to identify new customer targets Performed recency, frequency, and monetary analysis to identify new customer targets with excel Developed dashboard to present data summary and results of the analysis

Sorted consumer-transaction and demographic datasets in Excel, performed recency-frequency-monetary analysis to identify customer targets, and presented the methodology and findings in a dashboard and PowerPoint.

Outputted outstanding diner satisfaction experience while taking multiple roles in a fast-paced environment Received trust and positive feedbacks from both the original and new manager during the transition phase Built rapport with shift members while taking responsibility during both opening and closing shifts

Helped oversee a 25,000-square-foot dining operation serving up to 2,500 customers and reduced unnecessary overstay hours approximately one-third through tighter shift controls.

Collaborated with experienced organizers and performers for Chinese New Year Performance Scheduled and supervised fundraisers at various shops and restaurants for club events Created club’s poster designs, flyer designs, and advertisements for all club events

Conditioned students’ time management through practices of pacing and usage of alarms induced students to study advanced materials to meet above grade level expectations Instructed students’ English skills through additional translation and clarification

Provided bilingual tutoring, incorporated advanced material into student curricula, and built long-term client relationships through on-demand Mandarin-English instruction.