Roast My TechStack #4 · May 7, 2026 · Singapore
Three stacks on the table: a production-style personal LinkedIn analytics pipeline, a caregiver navigation app, and a voice-driven resume. This recap preserves the practical decisions, trade-offs, and questions raised in the room.
Speaker 01
Personal LinkedIn Analytics Pipeline
Ying Zhao
A production-style personal analytics pipeline for LinkedIn, built with Databricks in a medallion architecture — and optimized for maximum reach. The roast covered getting the data, ETL done right, retries, and the One Big Table trade-off.
Databricks · Medallion ETL · One Big Table · API · Retries & backoff
01
How To Get The Data
- API — the first-class route.
- Scraping — when the API doesn't give you what you need.
- 3rd party analytics — as a complement or fallback.
02
ETL In Medallion Terms
E → Extract (bronze) · T → Transformation (silver) · L → Load (gold). Databricks is the cloud-based platform that runs the data analytics and AI workloads.
03
Retries And Backoff
To overcome time-out issues, consider retries with backoff — especially for LLM calls and APIs. This prevents a huge bill shock when a provider starts failing in burst.
04
One Big Table Over A Relationship DB
The One Big Table approach beats a relationship database (e.g. Neo4j) for query speed and simplifies data consumption — fewer joins, one place to ask questions of.
Speaker 02
Care Compass
Ee One Lim
An app that navigates caregivers to the right support resources — a RAG-over-LLM build that raises reliability over raw generation. The roast dug into RAG vs. web search, dynamic model selection, and where to store the knowledge base.
RAG · n8n · Postgres · Dynamic model selector · Web search
01
RAG Over Plain LLM
If you have a single source of truth, RAG is preferred over a raw LLM — it increases reliability. And if you need the LLM to do a web search, use a dedicated web search tool rather than letting the model guess.
02
Dynamic Model Selection With n8n
Consider an n8n dynamic model selector to choose the best model per use case. n8n reduces LLM usage overall and can switch between using an LLM web search and RAG depending on the query.
03
Store The Knowledge Base In Postgres
Consider Postgres for the RAG store. For cost efficiency, consider other LLMs for this use case — like GPT-5.4 mini or other open-weight models.
Other tools raised: the eval tool in the Claude platform, and Kapso for WhatsApp business workflows.
Speaker 03
Voice Resume
Kaleb Nim
A voice-driven resume — an AI-voice interview app that lets employers ask anything and get real-time answers. The roast covered voice recognition, budget compute, text-to-speech, and guardrailing AI applications.
ASR · Whisper TTS · Oracle Cloud free tier · Guardrails · persona.md
01
Speech In, Speech Out
ASR (automatic speech recognition) turns the interview questions into text, and Whisper does the text-to-speech for the answers. The result: employers talk to a resume like it's a candidate.
02
Budget Compute That Lasts
For a budget option, Oracle Cloud's free tier lasts forever: 24GB RAM, 4 CPU, and 10TB egress — plenty for a side project with a video-heavy workload.
03
Guardrails For AI Applications
Always consider guardrails for AI applications, and the simplest way to build them is via a persona.md — encode who the AI is and what it will not do right in the prompt, checked before any answer ships.
Venue sponsor
SGInnovate
PIC: Arul · Co-hosted with GrowthLab
Event ops volunteers
⏳ Time & Mic Keeper
Esther
📸 Memories Keeper
Arul
🎙 Host
Jesse
🧠 Knowledge Keeper
Han