AI Software Engineer · Kerala, India

Aravind E S

I'm a curious, deliberate engineer who likes understanding how systems behave under pressure, then making them useful, dependable, and clear to the people who rely on them.

↓ Download resume View projects Chat with my AI

2 yrs

production AI

U.S. safety certified solution

flagship projects

Selected projects

Four systems that shipped into compliance critical production — plus an open-source Android TV app.

Agentic AI Inventory Intelligence Platform

2024–25

FastAPIMCPRAGPydantic

Production LLM agent for live railcar inventory with read only access, domain grounded RAG, and a skill discovery layer.

Architected a read only LLM agent with Pydantic schema validation, so no hallucinated writes reach production data.
Built RAG grounding agent reasoning in live database state and operating-procedure documentation.
Added Redis caching and async skill-discovery that learns reusable query patterns without added latency.
Engineered the sensor data layer (RFID, track-switch telegrams, PostgreSQL, Redis) for real-time inventory tracking.

ImpactShipped to the client for production use.

Real-Time Edge Perception & Collision Warning System

2026

ONNXTensorRTCUDAJetson

Real time edge perception and collision warning on NVIDIA Jetson, fusing camera, radar, GPS, and a track occupancy regression model.

Optimized TwinLiteNetPlus object detection through ONNX, TensorRT, and CUDA for low-latency Jetson inference.
Fused camera ego path detection with radar and GPS telemetry for robust risk analysis.
Trained and optimized a regression model for track occupancy and future railcar position (classical ML + deep learning).
Built multithreaded pipelines for embedded automotive compute, memory, and power limits.

ImpactHelped develop a collision warning solution that received U.S. safety certification for production deployment.

Agentic AI Clinical Fall Intelligence Platform

2025

FastAPIMCPLLM Tool CallingDocker

Clinical fall risk analysis agent where the model works through controlled tools, not raw patient data.

Built typed function-calling contracts (MCP) for safe clinical data access.
Kept model reasoning separate from sensitive records via deterministic APIs.
Containerized the backend on AWS EC2 with Nginx reverse proxying.

ImpactA production-style agent pattern for high-trust healthcare analytics.

Goal Progress & Rehabilitation Intelligence Platform

2026

PythonFastAPIPydanticLLMs

Rehabilitation progress intelligence with deterministic metrics and a separate LLM review layer.

Computed completion, velocity, and timeline deviation from structured activity logs, with no LLM in that path.
Used LLMs only to explain trusted metrics in plain language.
Validated the interface between analytics and language generation with Pydantic.

ImpactReadable progress summaries without letting the LLM rewrite the truth.

TGTV — Telegram as your private Netflix on Android TV

2026GitHub ↗

KotlinTDLibExoPlayerAndroid TV

Open-source Android TV app that streams videos from your own Telegram account like a private Netflix, with in-process TDLib range streaming and no third-party hosts.

Streams Telegram-hosted videos in-process: TDLib byte-range downloads read straight into ExoPlayer, no Bot API cap, no external hosts.
Two login modes: curator bot as a library appliance, or QR phone login for the full chat history.
TMDB metadata pipeline with release-name parsing, spelling-tolerant retries and web-assisted resolution.
Resume playback, OpenSubtitles fallback, storage-aware streaming and automatic cache hygiene (2 GB cap).

ImpactReleased v1.1 and in daily use on a Xiaomi Mi Box; source, README and release APKs are public on GitHub.

Experience

AI Software Engineer

Tata Elxsi · Kerala, India

Dec 2024 – Present

Architected a production LLM agent enforcing read-only access, validated queries, and Pydantic schemas — no hallucinated writes reach live railcar data.
Built domain-grounded RAG fusing real-time DB state with operating procedures; shipped Redis-cached async skill-discovery with zero user-facing latency.
Developed real-time perception inference pipelines with TensorRT, CUDA, and NVIDIA Jetson — ~25% lower inference latency; helped develop a U.S. safety-certified collision-warning solution.
Deployed CV workloads (YOLO, SAM 2.1, OCR/VLM, multi-camera tracking) on Jetson under tight compute/memory/power budgets; stabilized RTSP/FFMPEG/GStreamer streaming by ~30% under degraded networks.
Owned SQLite → PostgreSQL / Azure PostgreSQL migration for concurrent high-throughput production workloads.

LLM AgentsRAGTensorRTJetsonYOLOMCPSensor Fusion

AI Software Developer Intern

Tata Elxsi · Kerala, India

Jan 2024 – Jun 2024

Developed modified YOLO architectures with dual detection heads for extended detection on pretrained models.
Built radar point-cloud processing pipelines using ROS for real-time perception systems.
Developed a multi-camera streaming backend service with sockets and multithreading for real-time inference.

YOLOROSMulti-CameraRadar

Skills

Two years of focused production work. Grouped by domain.

AI / ML & Computer Vision

YOLO
SAM 2.1
TwinLiteNetPlus
OCR/VLM
OpenCV
Camera Calibration
Multi-Camera Tracking
Deep Learning
Model Optimization
QLoRA
MLflow

GenAI & Agentic AI

LLM Agents
LLM Engineering
RAG
Agentic Workflows
LLM Tool Calling
MCP
LangChain
LlamaIndex
Pydantic Schema Validation

Edge & Sensor Systems

NVIDIA Jetson
Edge Inference
ROS
Radar Point Clouds
Sensor Fusion
Low-Latency Pipeline
Multithreading

Inference Optimization

TensorRT
CUDA
ONNX
Model Optimization

Backend & APIs

Python
C++
FastAPI
Flask
REST APIs
Microservices
Pydantic
Celery

Databases & Caching

PostgreSQL
SQLite
Redis

Cloud & DevOps

Docker
Kubernetes
AWS EC2
AWS Bedrock
SageMaker
Azure
Google Cloud
Nginx
Git

Languages

Python
C++
C
SQL
TypeScript

Let's talk

Open to full time AI/ML roles. Fastest path: email or LinkedIn.

Open to opportunities

↓ Download resume