CP3405DT3 · System Intelligence Platform
Prof. Dr. Tan · TR2 '26
SPRINT W3 · ARCHITECTURE DEEP DIVE · From manual process to intelligent platform
ARCHITECTURE DEEP DIVE

From Manual Sprint to Intelligent Platform

A technical examination of the CP3405 system: how it works today, how it should work tomorrow, and what your team's manual work is actually specifying.

Diagram 5 — Predictive → Prescriptive Analytics (9 Stages)

From Data Flow to Calibrated Intelligence

This is not a market prediction exercise. It is a Scrum exercise. The market is the environment your sprint operates in. The prediction is your sprint increment. The delta is your Sprint Review. The prescription is your Retrospective action item. Every stage maps to a Scrum ceremony — watch for the label at the top of the canvas as you advance.

Click Next ▶ to step through each stage at your own pace. Audio narrates when you advance. Stages 1–8 are core. Stage 9 is a bonus extension for advanced teams.

STAGE:
STAGE 1/9
Click Next ▶ to begin. Audio will narrate each stage as you advance. You can also click any dot above to jump to a specific stage.
● PHASE 1 — Sprint Pipeline (Stages 1–5)
● PHASE 2 — Prescriptive Calibration (Stages 6–7)
● CORE — 11 Sectors (Stage 8)
★ BONUS — Multi-Timeframe (Stage 9)

Use Next ▶ to advance. ◀ Back to review. Click any stage dot to jump directly. 🔊 Audio plays when you advance a stage.

LLM Architecture

Four Models — One Identical Prompt — What Divergence Reveals

Querying four LLMs with identical inputs is not redundant. Divergence is signal. Agreement is confidence. Understanding why they differ is what separates an intelligent system from a chatbot wrapper.

ChatGPT (OpenAI)
Strong on narrative synthesis and macro framing. Watch for confident-sounding hallucinations on specific data. Best used for reasoning through scenarios.
Claude (Anthropic)
Strong on nuanced reasoning, uncertainty acknowledgement, and structured analysis. More likely to flag caveats. Good calibration signal for confidence scoring.
Gemini (Google)
Strong on recency and web-grounded data. May surface recent news others miss. Check for factual grounding and source quality in outputs.
DeepSeek
Different training distribution. Useful as a divergence check. When DeepSeek alone disagrees with the other three, investigate why — it may be picking up something different.
How to read the comparison table: When 4/4 models agree → high consensus signal (verify calibration). When 3/1 split → investigate the outlier. When 2/2 split → maximum uncertainty; Human Score R7 lead carries the decision.
From Manual to Automated

What Each Step Looks Like When Automated

Manual Now (Sprint W3)Automated in Prism (Future)Human Role Remaining
Screenshot Finviz sector mapAPI call fetches sector data on scheduleReview anomalies flagged by system
Write Almanac analysis by handCalendar parser + historical pattern DBApprove or override seasonal weight
Run 4 LLM prompts manuallyParallel API calls, auto-saved responsesReview comparison and set weight
Build comparison table in WordAuto-generated agreement matrixInterpret divergence, apply scepticism
R7 lead writes Human ScoreStructured input form with guided fieldsWild Card insight — irreplaceable
Create GitHub release manuallyCI/CD pipeline auto-tags on approvalFinal approval click
Post Discord submission by handBot posts from template on release triggerVerify accuracy before send
PRISM Platform Reference

Where This Is Heading — The Prism Architecture

Prism (Predictive Reasoning and Intelligence Synthesis for Markets) is Professor Dr. Tan's deployed AI market intelligence platform. It embodies the exact architecture you are learning to specify manually this sprint.

Flask + FinBERT Backend
Python Flask serving a FinBERT sentiment model. Market text → structured sentiment signal. This is your "Macro Agent R4" as code.
TradingView Lightweight Charts
Interactive candlestick rendering for SPX, NDX, IWM. Visual layer on top of the Technical Agent R5 output.
Multi-LLM Insights Panel
ChatGPT and DeepSeek panels running identical prompts. The comparison table your R8 operator builds manually — automated.

Architecture Guiding Questions — Sprint Retrospective Prompts

What was the most repetitive step this sprint?
That is your first automation candidate. Define the input, output, and frequency. Write it as a system requirement.
Where did the four LLMs disagree most?
Disagreement zones reveal where human judgement is most valuable and where the system should request escalation.
What would break if your team had 50 members?
Current processes that rely on one person knowing something informally are not scalable. Document and systematise them.
What data are you not collecting yet?
Every Almanac, Macro, and Technical output implies a data source. Which ones are missing from your current pipeline?
← Back to Sprint Instructions System Architecture Overview