Aliya A

Associate AI Consultant

Technoloy • Período integral

Sep 2024 - May 2026 • 1 yr 8 mos

Built and deployed a production agentic news intelligence pipeline ingesting 467 live multi-region sources daily for a Malaysian federal government client, working directly with government stakeholders to define ingestion and summarization requirements. Mitigated AWS Bedrock context-window and token-limit failures by building an HTML tokenization and structural chunking layer that stripped DOM boilerplate and cut per-query input from 12k to 7.8k tokens (35% reduction). Paired this with model tiering, routing translation to Claude Haiku and summarization to Sonnet, to hold inference cost within the daily government budget. Conducted a formal multi-agent architecture evaluation, supervisor-agent pattern vs. LangGraph node-state machine, for a large-scale government LLM application. Recommended and implemented LangGraph for deterministic routing, zero per-step LLM overhead, and elimination of infinite-loop risk, dropping P99 latency from 4.2s to 3.0s and removing state-concurrency errors. Led a 2-person build of a multilingual RAG chatbot for a Malaysian federal government client across English, Bahasa Malaysia, Tamil, and Mandarin, on AWS Bedrock with Cohere multilingual embeddings and an OpenSearch vector store. Added real-time voice input via AWS Transcribe streaming and a Next.js and TypeScript front-end. Built and deployed an enterprise customer support AI agent integrating LangFlow, Chatwoot, and AWS Comprehend sentiment analysis. Programmatic routing rules offload dissatisfied users to human agents in real time based on negative-sentiment thresholds, removing manual sentiment triage. Sole AI engineer for the consultancy's internal LLM tooling through a firm-wide product pivot, migrating legacy pipelines toward modular backend services.

AI Engineer

Tech Cloud • Período integral

Jul 2023 - Aug 2024 • 1 yr 1 mo

Designed and deployed an internal HR RAG chatbot enabling semantic search over structured and unstructured HR documents, establishing auditable retrieval workflows that fully replaced the manual HR query resolution process for the consultancy team. Built a WhatsApp-based AI tutor with multi-turn context management and adaptive prompt routing, serving 200+ concurrent users across bootcamp and professional cohorts. Owned the end-to-end ML pipeline from raw satellite imagery ingestion to containerized production inference across 20,000+ global maritime port locations on AWS. Trained YOLOv8-OBB (oriented bounding boxes) for vessel detection at 0.84 mAP, a 15% improvement over baseline, through custom anchor tuning for vessel aspect ratios and Mosaic data augmentation. Built a spatio-temporal adjacent-frame extraction technique to recover occluded marina zones from temporally neighboring satellite frames, improving port coverage completeness by 25% and reaching full coverage on previously-occluded zones Built a cashierless checkout proof-of-concept integrating YOLOv8 with DeepSORT multi-object tracking. Independently curated, labelled, and balanced a 2,500-image dataset, achieving 97% real-time tracking accuracy with minimal ID-switching. Authored a 20-page data annotation guide adopted as the engineering team's standard for multi-class industrial datasets, establishing a unified ground-truth standard and reducing labeling errors by about 40%. Upgraded legacy computer vision pipelines, including ANPR and industrial rubber defect detection, from YOLOv5 to YOLOv8, improving inference throughput and mAP for edge-constrained deployment. Fine-tuned object detection and multimodal OCR models for high-variance industrial use cases: multi-ingredient recognition for a commercial pizza chain, durian maturity classification (empty/half/full), port container OCR, and rubber defect detection in manufacturing pipelines.