Igniters - [Week3] Production Track: Alex - Agentic Equities Explainer

[iJoshy]

Alex - Agentic Learning Equities Explainer

Enterprise-grade, multi-agent financial planning SaaS built on AWS + OpenAI Agents SDK
Researches markets and ingests knowledge

Architecture

graph TB
    Researcher[Researcher Agent on App Runner] --> Bedrock[AWS Bedrock<br/>gpt-oss-120b]
    Researcher --> Ingest[Ingest Lambda + API Gateway]
    Ingest --> SageMaker[SageMaker Embeddings]
    Ingest --> V[(S3 Vectors)]

Loading

Improvements Added in This Implementation

The following enhancements were added to make the researcher significantly more agentic and resilient:

Area	Improvement	Why It Matters
Context Engineering	Introduced ResearchBrief and render_research_task with explicit objective, guardrails, output sections, and quality checks	Better consistency and less drift in agent behavior
Agentic Planning	Added run-scoped to-do tools: add_todo_item, complete_todo_item, get_todo_status	Forces explicit plan/execute/verify loop
Source Discipline	Added record_source and get_source_log tools	Creates a traceable evidence ledger per run
Ingestion Safety	Added per-run ingestion guard (ingestion_count) to allow ingest once	Prevents duplicate writes and noisy vector store entries
Model/Region Config	Replaced hardcoded model/region with env + defaults (gpt-oss-120b, us-west-2)	Safer multi-env deployment and easier tuning
Terraform Runtime Env	Added bedrock_model_id + bedrock_region variables and App Runner env wiring	Infrastructure-driven model control
Deploy Script Hardening	Updated deploy pipeline to pass Bedrock env vars during App Runner update	Reduces config drift after redeploys
Timeout Resilience	Added bounded runtime controls (RESEARCHER_MCP_TIMEOUT_SECONDS, RESEARCHER_MAX_TURNS, RESEARCHER_REQUEST_TIMEOUT_SECONDS)	Prevents runaway runs and App Runner 504s
Fallback Path	Added automatic fallback agent run when MCP/turn/time limits fail	Endpoint returns useful output instead of failing hard
Prompt Guardrails	Explicitly disallowed unavailable browser tools (e.g. browser_search) and tightened source limits	Reduces tool mismatch errors and improves reliability

Files Updated for These Improvements

• backend/researcher/context.py
• backend/researcher/tools.py
• backend/researcher/server.py
• backend/researcher/deploy.py
• terraform/4_researcher/variables.tf
• terraform/4_researcher/main.tf
• terraform/4_researcher/terraform.tfvars.example

Researcher Runtime Tuning

These env vars control reliability/latency behavior in backend/researcher/server.py:

BEDROCK_MODEL_ID=openai.gpt-oss-120b-1:0
BEDROCK_REGION=eu-west-1

RESEARCHER_MCP_TIMEOUT_SECONDS=30
RESEARCHER_MAX_TURNS=14
RESEARCHER_REQUEST_TIMEOUT_SECONDS=75

Tune recommendations:

• Lower values: faster responses, more fallback usage
• Higher values: deeper browsing, higher timeout risk

Roadmap Ideas

• Add Polygon MCP integration for market-price enrichment
• Add structured “fact table” extraction pipeline to external DB (e.g., Supabase)
• Add SQS retry wrappers around researcher-triggered ingestion
• Add evaluator/critic agent for report quality scoring
• Add user-facing citation cards sourced from record_source ledger

Application Source

The complete application is available at: https://github.com/iJoshy/alex

See the README.md in the repository for detailed information about implementation, setup, and usage.

Other Week 3 Exercises

🚀 Cybersecurity Analyst on AZURE
🚀 Cybersecurity Analyst on GCP

@iamumarjaved
Cc @ed-donner

[iamumarjaved]

Great to see @iJoshy Good efforts

@ed-donner This is ready to be merged.