AI/ML · Flex — SCU Practicum
Flex Competitive Intelligence Platform
AI platform that turns 15+ hours of weekly competitor research into 2-hour automated intelligence briefs
The Challenge
Flex's strategy team manually monitors 4 competitors across SEC filings, earnings calls, patents, and hiring trends — spending 15+ hours per week. Insights arrive stale, live in scattered spreadsheets, and lack the synthesis needed for CapEx investment decisions. No centralized platform exists for competitive intelligence in the contract manufacturing sector.
My Role & Approach
Defined the product vision, 12-week roadmap, and weekly milestones for a 4-person team. Owned the RAG pipeline architecture, API design, and auto-ingestion scheduler. Made key decisions on cost constraints ($20-50/month ceiling), hybrid search strategy, and data source prioritization. Coordinated weekly demos and managed API-first development process.
Product Thinking
Problem Framing
The initial brief was 'build a competitor analysis tool.' Through stakeholder interviews with Flex's Strategy and Sales teams, we discovered the real problem: analysts weren't lacking data — they were drowning in it. The actual need was synthesis, not search. This reframing changed our product from a document search engine to a conversational intelligence platform.
Key Tradeoffs
Hybrid RAG + web search vs. RAG-only: RAG alone returned stale data (filings are quarterly), web search alone lacked proprietary context. We chose the harder path of combining both with source attribution. Also chose Claude over GPT-4 for cost — at $20-50/month total budget, every API call mattered.
What We Didn't Build
Cut automated email alerts, competitor SWOT generator, and PowerPoint export. The team wanted all 12 features in 12 weeks. I led the decision to ship core conversational RAG first and validate with users before building more. Three features were permanently cut after user testing showed they weren't needed.
Cross-Functional Leadership
Leading a 4-person team across frontend, backend, and data engineering. Running weekly demos for Flex stakeholders in Sales, Strategy, and Executive teams. Managing sprint planning, backlog prioritization, and API-first coordination between team members.
Solution & Execution
Full-stack AI platform with automated data ingestion from SEC EDGAR (daily at 4PM ET), earnings transcripts, patent filings, job boards, and news RSS. ChromaDB vector store for semantic search across ~405 documents. Hybrid RAG + Brave web search pipeline powered by Claude API — neither RAG alone (stale) nor web search alone (no proprietary context) was sufficient. Next.js dashboard with conversation memory, sentiment tracking, and anomaly detection for CapEx spikes. Total infrastructure cost: ~$20-50/month.
Tech: ChromaDB, Claude API, FastAPI, Next.js, Python, Brave Search, Sentence Transformers
Impact
- ~405 documents indexed across 5 target companies (Flex, Jabil, Celestica, Benchmark, Sanmina)
- Hybrid RAG + web search with conversation memory for natural follow-up queries
- Auto-ingestion of SEC filings (daily), patents (weekly), job postings (daily)
- ~$20-50/month total cost — all free-tier infrastructure with Claude API as only paid service
- (In progress — user testing and metric collection ongoing)
Challenges & Pivots
Challenge: Scope creep — team wanted to build 12 features in 12 weeks
Resolution: Cut 3 planned features to ship core RAG chat first. Established 'working demo every Friday' rule to keep the team focused on shippable increments.
Challenge: Financial documents (10-Ks) broke with naive text chunking — tables and CapEx figures lost context
Resolution: Implemented domain-aware chunking that preserves table structures and keeps related financial metrics in the same chunk. Improved retrieval relevance significantly for CapEx queries.
What I Learned
- Leading a 4-person team requires ruthless scope prioritization — the hardest PM skill is saying 'no' to good ideas.
- API-first development (define endpoints before building) eliminated most frontend-backend integration issues.
- Hybrid search (RAG + web) is significantly more useful than either alone for competitive intelligence workflows.
Next case study
Nazava × Shopee Sales Optimization