Deep Dive

Tell me about your Atlas project

• manual DoD procurement → slow, error-prone

• goal: automate + standardize

• AI-powered system

  • validation (dod procurement forms)

  • chatbot (form help + Q&A)

  • form generation assist

• impact: 97% faster processing

What was the architecture? - Atlas Project

• frontend: Next.js + Tailwind

• backend: FastAPI (async APIs)

• DB: PostgreSQL (JSONB)

• AI: AWS Bedrock (LLM)

• infra: AWS (S3, Lambda), Docker/K8s

• flow: user → API → AI + DB → UI

How did you design the frontend? - Atlas Project

• reusable components

• forms + validation UI

• separation of concerns

• scalable structure

How did you manage state? - Atlas Project

• local + lifted state

• shared → centralized

• avoided prop drilling

• Redux if scale increases

How does data flow? - Atlas Project

Chat path: User message → JWT auth → intent detection (informational vs project toggle) → prompt built → Claude 3.5 Haiku on Bedrock → response comes back. If project mode, AI extracts NTB form fields from the response — high confidence (≥0.95) auto-confirms, medium (0.70-0.94) goes to a clarification queue for user verification, low gets tossed. Once fields are confirmed, user can generate a PDF.

Document path: PDF uploaded → S3 → Textract parses it → rule-based checklist + AI validation → stored in PostgreSQL.

Vendor path: Admin creates projects with deliverables → vendors upload files → admin approves/rejects.

How do you keep it fast?

- Atlas Project

• async AI calls

• OpenSearch (if mentioned)

• pagination

• avoid blocking UI

Biggest challenges? - Atlas Project

• AI inconsistency → validation layer

• large datasets → UI latency

• async workflows

What tradeoffs did you make? - Atlas Project

• no WebSockets → not needed

• simple state > Redux early

• avoid over-engineering

What breaks at scale? - Atlas Project

• FE: rendering large lists

• BE: query performance

• AI latency

• fix: pagination, indexing, async

What would you improve? - Atlas

• Redux if complexity grows

• optimistic UI

• better caching

Why Next.js? - Atlas Project

• SSR → performance

• full-stack integration

• routing simplicity

Why not Redux? - Atlas Project

• overkill at current scale

• added complexity

• introduce when needed

Prevent re-renders? - Atlas Project

• memoization

• component isolation

• pagination

Handling large data? - Atlas Project

• pagination

• lazy loading

• avoid full renders

Latency sources? - Atlas Project

• network

• DB queries

• AI processing

How does the AI system work? - Atlas Project

• chatbot (intent detection)

• validation (LLM + rules)

• form generation (structured extraction)

• confidence scoring

What does the system do?

Atlas Project

• chatbot → natural input

• form generation (7600A, MIPR, NTB)

• PDF generation

• admin workflow (submit → approve → reject)

How is the system secured?

Atlas Project

• Keycloak (auth)

• RBAC

• JWT

• least privilege

• encryption (S3/KMS)

How does the system scale?

Atlas Project

• async APIs (FastAPI)

• AWS services

• stateless backend

• horizontal scaling

Who are the users?

Atlas Project

• admin (review/approve)

• customer (submit forms)

• vendor (track deliverables)

What makes this system unique? - Atlas Project

• natural language → structured forms

• AI validation (not just generation)

• end-to-end automation

• replaces manual workflow

Engineering Narrative Framework

👉 This is your actual structure:

1. Problem

2. System / Architecture

3. Your Contributions (focus on frontend)

4. Challenges

5. Tradeoffs

6. Impact

7. Improvements