AI Engineer at Heu.ai · Kerala, India
I build practical AI systems: the work around the model. Reliable APIs, retrieval, evaluation and data pipelines, in Python and Rust. I like turning ambiguous problems into software people can inspect, test and use.
Now: building applied AI systems with reliable retrieval, evaluation and backend workflows.
Crisis-Graph: flood emergency dispatch where the LLM extracts and deterministic code decides. Rust services turn multilingual SOS messages (English, Malayalam, Manglish) into verified vehicle dispatches. A fine-tuned Qwen3-4B only proposes a candidate report; routing, allocation and per-segment verification make every decision, and anything uncertain goes to a person.
- 40 → 81 exact extractions on 120 blind cases after QLoRA fine-tuning
- 20 reports/s surge with every report accepted and 0 unverified routes
- 12 → 39/min T4 throughput after fixing a grammar bottleneck, identical answers
Rust Tokio Axum Neo4j Redis Streams Qwen3-4B QLoRA vLLM llama.cpp OpenTelemetry
EcoSwapChain (college project, 2025): a recommerce platform where every item is an NFT and every sale goes through escrow. Django API and workers · React app · Solana service · sustainability model
Action Precision Analyzer (college project, 2024–25): scores athletic movement from video. MediaPipe pose landmarks → joint angles → a two-headed LSTM.
Full case studies: amal-antony-dev.vercel.app
| AI/ML & agents | PyTorch · Transformers · PEFT/QLoRA · LangGraph · LangChain · Ragas · RAG · vLLM |
| Languages & backend | Python · Rust · SQL · JavaScript · C++ · FastAPI · Django · Flask · Celery |
| Data & messaging | PostgreSQL · DynamoDB · Redis Streams · Neo4j · Qdrant · Pinecone · RabbitMQ |
| Cloud & observability | AWS · GCP · Prometheus · Grafana · OpenTelemetry · LangSmith |