Swaraj Shaw
Machine Learning Engineer · NLP · Speech · LLM Systems
Dublin, Ireland · shaw.swaraj16@gmail.com · +353 89 984 9430
Summary
Machine Learning Engineer with 5+ years of experience delivering NLP, Speech (TTS/ASR), and LLM-based systems across Big Tech and startups. Expert in text-to-speech modelling, embeddings, ONNX inference, Python/Rust ML infrastructure, evaluation design, and large-scale data quality automation. Strong track record building high-impact AI platforms, offline inference engines, ML pipelines, and production-grade products. Currently building full-stack Irish mobility platform AutoHub Ireland and pursuing an LLM Agentic AI course at DkIT while working at Meta.
Technical Skills
Experience
- SME for Hindi TTS: phoneset design, G2P mappings, TN rules, and full linguistic pipeline.
- Built a 14k+ sentence evaluation suite across 11 linguistic categories; automated coverage analysis.
- Corrected 14k+ phonetic errors in a 35k-word lexicon database using Python tooling.
- Achieved 95.6% benchmark accuracy through iterative training and error analysis.
- Major contributor to Ray-Ban Meta smart glasses Hindi TTS: evaluated 30k+ scripts, shortlisted 200+ voices.
- Collaborated with engineering on training platforms (Bento, MLHub) and quality tooling.
- Built conversational chatbots (Amazon Lex + Connect) across Slack, Messenger, and internal tools.
- Delivered a recommendation engine via Amazon Personalize + SageMaker, increasing conversions by 15%.
- Built anomaly detection pipelines for energy data; reduced operational wastage by 15%.
- Automated BI dashboards (QuickSight, Tableau); improved data visibility and onboarding KPIs by 50%.
- Built credit-risk and churn models using SMOTE + XGBoost; improved accuracy by 15%.
- Prototyped ASR-based booking and customer support workflows integrated with ML APIs.
- Built collaborative filtering recommendation models improving sales by 23%.
- Created CNN-based image classifiers achieving 95% accuracy for internal automation.
- Led Android app for Quest Mall; achieved 5,000+ installs and increased footfall via location-intelligent features.
Projects
Irish-localised vehicle intelligence platform: real-time fuel & EV price map, best-value routing, RSA driving test scraper, and an AI Master Mechanic assistant with RAG over vehicle manuals.
- React Native mobile app (Expo) + Next.js web portal with Supabase auth
- NestJS API + PostgreSQL monorepo with Docker + Render deployment
- Playwright-based automated RSA wait-time scraper running on cron
- Community features: verified owner badges, contributions, achievements
Enterprise AI orchestration monorepo: FastAPI gateway with prompt orchestration, Next.js 19 messenger UI, async Celery workers, QLoRA LLM fine-tuning, and Whisper ASR adaptation.
- Knowledge Distillation: Automated GPT-4o teacher to local GGUF/student distillation pipeline.
- Compliance ETL: PII-redaction-aware dataset exports (spaCy-powered) with audit trails.
- Unified MLOps: Prefect-orchestrated nightly training, S3 model registry, and Slack-integrated failure reporting.
- AI Lab: Full-stack evaluation suite for real-time model comparison using BLEU/WER metrics.
Local-first AI file intelligence desktop app. Rust daemon with GPU-accelerated ONNX inference (Metal/DirectML) semantically renames, deduplicates, and organises files — all fully offline.
- MiniLM/BGE/Gemma embeddings for semantic rename & folder recommendations
- SQLite (WAL) metadata + audit trail — every action undo-safe
- Tauri UI with previews, batch actions, confidence threshold config
- C++ ONNX Runtime inference engine with Metal + DirectML GPU backends
Run Llama 3, Mistral, and Phi-3 on any laptop — no GPU, no cloud, no login. Uses libp2p DHT to split model layers across peers, like BitTorrent for LLM inference.
- libp2p DHT peer discovery: layers distributed across the network
- Encrypted prompts — no server, no IP logging, fully auditable
- Supports any Hugging Face GGUF model ID out of the box
Proprietary hiring intelligence platform and automation cockpit. Features custom scrapers, heuristic/ML detail normalization, and enterprise-grade recruitment signals orchestration.
- MiniLM-powered JD field extraction with automated retraining loop (Corrections → JSONL → Embeddings → Model)
- High-performance Rust-based HTML extractors and job feed parsers for native-speed ingestion
- Modular Node.js + PostgreSQL backend with invite-only access, admin impersonation, and job-board localization
- Signal-rich analytics dashboard for live recruitment automation orchestration
Virtual fencing platform for cattle using AirTags and BLE devices — a cost-effective alternative to Nofence. Farmers draw boundaries on a mobile map; smart alerts fire when animals approach fence lines.
- Cross-platform mobile app with real-time location tracking and alerts
- AI behaviour analysis to distinguish grazing from fence-breaking events
- Universal device support: AirTags, Tile, and custom collar integration
Agentic AI music assistant built for DkIT LLM course. Implements both ReAct loop and Workflow patterns with Pydantic tool-calling (album lookup, artist search, genre queries) over a SQLite music database.
- ReAct agent: single prompt → tool call loop → LLM summary
- Workflow mode: CSV batch processing, one enquiry at a time
- LangSmith tracing + Groq inference backend
Year-long production AI engineering curriculum at Dundalk Institute of Technology. Building enterprise-grade systems with RAG, agents, evals in CI, security controls, observability, and cost controls.
- Weekly labs: LangChain, vector DBs, agent workflows, prompt engineering
- Infra: Docker, Terraform, Kubernetes deployment pipelines
- Capstone: shippable, evaluated AI system with full observability
Full-stack accommodation management portal with web and mobile clients. Designed to simplify rental management for landlords and tenants, with real-time alerts and document management.
CNN-based skin condition classifier comparing ResNet, AlexNet, and VGG16 architectures under identical operational conditions. Best model deployed as a web application for real-time inference.
- ResNet, AlexNet, VGG16 evaluated with consistent train/val splits
- Flask web app deployment for live image classification
Automated document intelligence pipeline using AWS Textract + PyTorch for invoice and document parsing. Reduced manual document processing workload by 25% in production.
- AWS Textract for OCR with PyTorch post-processing for field extraction
- Structured output pipeline for invoices, forms, and receipts
Hybrid collaborative and content-based recommender system. Combines MovieLens dataset interactions with NLP-based content analysis to provide highly accurate, dual-signal movie suggestions.
- Hybrid engine: blends collaborative filtering with TF-IDF content similarity
- Advanced NLP: uses textual metadata and tags for semantic item matching
- Extensive evaluation: tested on ml-latest dataset with cold-start mitigation
Education
Let's build something together
Open to ML engineering roles, consulting, and research collaborations.