Roast My TechStack #6 · June 4, 2026 · Singapore
Two builds took the heat at SGInnovate: a parallel AI content pipeline and an evaluation harness for AI agents. This recap preserves the practical decisions, trade-offs, and questions raised in the room.
Speaker 01
AI Content Pipeline
Jesse Sng
Brought a content pipeline that fans one piece of source material out into a podcast, blog, images, and audio in parallel. The conversation concentrated on orchestrating parallel generation, engineering error handling so effort is never discarded, and choosing image models when text-in-image matters.
Parallel generation · Fail-continue · Imagen · GPT Image · Remotion · ComfyUI
01
Parallel Generation
When you run a parallel workflow — one source, multiple outputs at once — always plan for orchestration and piecing things together up front. The fan-out is the easy part; reassembling coherent pieces from independent branches is where the complexity lives.
02
Error Management: Continue First, Retry Later
If a branch fails, plan the workflow to continue first and go back to retry afterwards. Do not discard the previous effort — partial success is still worth keeping. Fail-continue beats fail-fast when each branch is expensive to produce.
03
Image Generation
Imagen is cheaper for image generation than GPT Image, but it does not support rendering text inside the image. If you need legible text baked into the asset, GPT Image remains affordable — or render and superimpose text over the image afterwards to save cost.
For finer control over image aesthetics, explore ComfyUI for image improvement pipelines. For programmatic video, consider Remotion.
04
Observability
Start light with Grafana for monitoring; move to Prometheus when the load warrants the heavier machinery.
Speaker 02
Evals for AI Agents
Gaurav Kumar Manek
A talk on evaluation for AI agents — where evals overlap with guardrails, how to bound an agent's blast radius, and when a deterministic check beats a model call. The clearest framework of the night for grading agent output.
LLM evals · Guardrails · Deterministic checks · Blast radius
01
Evals vs. Guardrails
Evals and guardrails can overlap, but they are not the same job. A guardrail can be implemented as deterministic code — it doesn't need to be another model.
02
Limit The Blast Radius
Two ways to keep a misbehaving agent from doing damage:
- By infrastructure — run agents on isolated resources (a separate Mac mini, a separate cloud account) so a failure can't reach production systems.
- By action — make destructive actions progressive and reversible: draft an email instead of sending it, mark an email instead of deleting it.
Venue sponsor
SGInnovate
PIC: An Ting
Event ops volunteers
⏳ Time & Mic Keeper
Jasmine
📸 Memories Keeper
Esther
🎙 Host
Han
🧠 Knowledge Keeper
Han