✅ MANUELA — COMPLETE TECHNICAL ASSESSMENT

Senior AI/ML + Blockchain + Cloud Architect.

Dataset used: US Border Crossing Port Activity Data (stored in MySQL us_border_crossing_data.border_crossing_entry_data).

Sample rows (same schema as CSV)

Port Name | State | Port Code | Border | Date | Measure | Value | Latitude | Longitude | Point Jackman | Maine | 0104 | US-Canada Border | Jan-24 | Trucks | 6556 | 45.806 | -70.397 | POINT (-70.396722 45.805661) Porthill | Idaho | 3308 | US-Canada Border | Apr-24 | Trucks | 98 | 49 | -116.499 | POINT (-116.49925 48.999861) San Luis | Arizona | 2608 | US-Mexico Border | Apr-24 | Buses | 10 | 32.485 | -114.782 | POINT (-114.7822222 32.485) Warroad | Minnesota | 3423 | US-Canada Border | Jan-24 | Personal Vehicle Passengers| 9266 | 48.999 | -95.377 | POINT (-95.376555 48.999) Ysleta | Texas | 2401 | US-Mexico Border | Jan-24 | Personal Vehicle Passengers| 521714 | 31.673 | -106.335 | POINT (-106.335449846028 31.6731261376859)

Reference this table directly when prepping data for QLoRA/RAG; Point is stored as a spatial POINT.

SECTION 1 — AI/ML + MODERN LLM ENGINEERING

1.1 Practical Conceptual Questions (Short Technical Answers Required)

These map directly to real production AI work.

Compare pretraining, SFT, RLHF, DPO, LoRA fine-tuning in LLMs.: Provide concise, technical differences and when to use each.
Explain how RAG reduces hallucinations when answering questions about border-crossing activity.: Be precise about retrieval, grounding, and context-aware scoring.
Explain tokenizers, positional encodings, and attention mechanisms using the border dataset as input examples.: Map how each component processes tabular text into model-ready tensors.
Explain KV-cache, quantization (4bit/8bit), speculative decoding and how they speed up inference for your RAG agent.: Highlight memory reuse, reduced precision math, and shorter search paths.
Compare vLLM vs TensorRT-LLM vs HuggingFace Transformers for high-throughput border analytics, small-batch low-latency inference, GPU optimization.: List criteria per workload and GPU tuning differences.
Explain how to protect your LLM API from prompt injection, jailbreak attempts, training data extraction (border records).: Describe guardrails, filters, and watermarking.

Each answer must be short (5–7 lines), highly technical, and precise.

1.2 Hands-On ML Task

Task: Build a Border-Crossing QA Model (QLoRA + RAG + API)

A. QLoRA Fine-Tuning (Llama-3-8B)

Deliverable: train_llama_border_qlora.py
Load Llama-3-8B using 4-bit quantization.
Apply QLoRA/PEFT adapters.
Convert your table into instruction-style JSONL.
Train using PyTorch + Hugging Face Trainer.
Save the LoRA adapter and log metrics (loss, perplexity).

B. Evaluation Pipeline

Deliverable: evaluate_border_qa.py
Implement Exact Match (EM), F1 score, ROUGE-L.
Use a small eval set manually derived from the dataset.

C. RAG Pipeline (FAISS)

Deliverable: rag_border_pipeline.py
Chunking (each row = one document).
Embedding using bge-small-en or gte-large.
FAISS index, retriever returning top-k rows.
Re-ranking with a Cross-Encoder.

D. Inference API (FastAPI)

Deliverable: app.py
Single endpoint /ask performing retrieval, re-rank, LLM generation, guardrails, and response with sources.

SECTION 2 — AI/ML ENGINEERING (PRODUCTION)

2.1 Architecture Challenge — Real-Time LLM Microservice

Design a production inference system for border analytics covering:

Support for vLLM or TensorRT-LLM.
GPU autoscaling with AWS EKS/Karpenter.
Server-Sent Events or WebSockets for streaming tokens.
Feature store for caching, user embeddings, safety scores.
Canary deployments.
Vector DB integration (FAISS / Pinecone).
A/B testing for different LoRA models.

Deliverables: Architecture Diagram, Component Responsibilities, Scaling Strategy, Failure Modes & Mitigations.

SECTION 3 — BLOCKCHAIN ENGINEERING

Use the border-crossing data as the economic asset: every record becomes an on-chain data point, users stake tokens on predictions, rewards depend on accuracy.

3.1 Smart Contract Assignment

Deliverables: /contracts/BorderStaking.sol and /test/border_staking.test.js or .t.sol.
ERC-20 staking with fixed APR and early withdrawal penalty.
Upgradeable via UUPS or Transparent Proxy.
Full Hardhat or Foundry test suite with gas optimizations and security protections.
Event logging for real-time analytics.

3.2 DApp Frontend

Deliverable: /frontend/ built with React + Vite + Wagmi.
Dashboard linking wallet (Wagmi + Ethers.js) with stake/withdraw UI.
Display user positions, APR, TVL, on-chain events.
Backend indexer using Node.js + PostgreSQL / DynamoDB at /indexer/.

SECTION 4 — BACKEND ENGINEERING (REST + GraphQL)

Deliverable: /backend/ supporting REST endpoints, GraphQL schema (borderSearch, nearestPorts, statistics), and JWT authentication.
Integrations with the LLM inference microservice, blockchain indexer, and AWS S3 dataset storage.
Provide Jest unit tests and a CI/CD pipeline (GitHub Actions or Bitbucket).

SECTION 5 — CLOUD + DEVOPS (AWS)

Deploy full system on AWS with ECS or EKS GPU nodes, DynamoDB or RDS for border records, SQS/SNS for async processing.
Expose APIs through API Gateway and serve the DApp via CloudFront.
Implement IAM least-privilege, Terraform or AWS CDK IaC, CloudWatch logging, and Grafana dashboards.
Deliverables: /infra/terraform/*, infrastructure diagram, and blue/green vs rolling deployment explanation.

SECTION 6 — FULL-STACK FRONTEND

Deliverable: /dashboard/ with real-time inference logs, GPU utilization (metrics API), border traffic visualizations.
Use Recharts or ECharts graphs with Zustand or Redux Toolkit state + Tailwind or Material UI styling.

SECTION 7 — FINAL SYSTEM DESIGN CHALLENGE

Design a full real-time AI-powered border analytics platform that:

Ingests border crossing data, streams events through Kafka.
Enriches + embeds documents, stores in a vector DB.
Supports AI assistant queries, 50k concurrent users, <1s inference, horizontally sharded historical data, GPU autoscaling, security & governance.

Deliverables: System sequence diagram, data flow + storage schema, API service layout, scaling strategy, and security considerations.