Production AI
Generative AI, RAG, agents, NLP-to-SQL, fine-tuning, evaluation, and real-time inference.
AI Engineer | Chennai, India
I build production AI systems using agents, LLMs, NLP, computer vision, and deep learning. My focus is simple: good accuracy, low latency, and reliable performance.
0%+
Agent orchestration
0%
NLP-to-SQL accuracy
<0.0s
Real-time AI target
0+
Production APIs
AGENTIC AILLM SYSTEMSCOMPUTER VISIONREAL-TIME INFERENCEAWS AI/MLDEEP LEARNINGAGENTIC AILLM SYSTEMSCOMPUTER VISIONREAL-TIME INFERENCEAWS AI/MLDEEP LEARNING
Profile
Engineering AI that performs outside the demo.
I connect applied AI research with disciplined software engineering to deliver systems that are accurate, observable, scalable, and useful in production.
End-to-end engineering
From training pipelines and APIs to event-driven AWS architecture, Kubernetes, and observability.
Measured outcomes
Accuracy, p90 latency, concurrency, reliability, and cost are treated as core product requirements.
Technical ownership
Led teams of 10+ members, completed 40+ code reviews, and contributed across design, delivery, and deployment.
Professional experience
AI Engineer | 1.5+ years
Designed production RAG systems, intelligent tool-routing agents, memory-enabled workflows, and real-time inference services.
Built 100+ production APIs following software-engineering standards and integrated AI capabilities into new and existing applications.
Optimized orchestration accuracy, NLP-to-SQL, retrieval quality, model inference, concurrency, and end-to-end latency.
Migrated deep-learning training and inference workloads from NVIDIA GPU infrastructure to AWS Trainium and Inferentia.
Worked across collaboration, audio generation, solar compliance, multilingual translation, and manufacturing quality domains.
Led 10+ team members, completed 40+ code reviews, and took ownership across architecture, implementation, optimization, and deployment.
Selected work
Complex AI systems, made operational.
Five production-focused engagements spanning agents, deep learning infrastructure, computer vision, multilingual inference, and industrial automation.
PROJECT 01
RedeAPP
Memory-enabled enterprise messaging intelligence
An agentic assistant for field workers that retrieves project, people, meeting, outing, and date-based information from PostgreSQL messaging data with low latency.
StrandsClaude Sonnet 4.5BedrockAgentCorePostgreSQLAWS Knowledge Bases
01Strands orchestration routes semantic questions to AWS Knowledge Bases and date/data questions to an NLP-to-SQL tool.
02A scheduled two-day preprocessing pipeline embeds recent data with Amazon Titan Embeddings V1 and stores vectors through AWS Knowledge Bases.
03AgentCore Memory provides long- and short-term conversation memory; AgentCore Runtime hosts the production agent.
95%+ orchestration accuracy
98% NLP-to-SQL accuracy
100% RAG accuracy
<1,200 ms at 50 concurrent users
PROJECT 02
SoundryAI
Audio generation on AWS Trainium & Inferentia
Migration of a multi-model audio-generation training and inference stack from CUDA/B200 GPUs to AWS purpose-built AI accelerators.
AWS NeuronTrainiumInferentiaEKSPyTorchHTDemucsFlowVAERVQ
01Migrated training to trn2.48xlarge and inference to inf2.24xlarge, using tensor parallelism across NeuronCores to resolve memory constraints.
02Stabilized training with memory optimization and gradient clipping, then built a UI to configure and launch training jobs.
03Deployed inference on EKS with HTDemucs, an autoregressive generator, RVQ tokenization, and FlowVAE decoding; PostgreSQL notifications queue requests for a five-user workload.
~2 min for a 3 min song
TP across 8 NeuronCores
Trainium training
Inferentia inference
PROJECT 03
Eequals Solar QC
Automated solar-installation compliance
An AI quality-control platform that evaluates installation photos, identifies defects, references compliance knowledge, and returns image-level pass/fail guidance.
RF-DETROCRLangChainLangGraphRAGAWS
01Fine-tuned RF-DETR to detect installation defects and combined it with OCR for labels and component evidence.
02Used LangChain for agent tooling and LangGraph for routing across visual analysis and a modifiable compliance RAG knowledge base.
03Designed low-confidence escalation, partial re-evaluation, admin-controlled knowledge updates, and AWS environments for Dev, UAT, and Production.
Seconds instead of hours
Per-image verdicts
Human escalation path
Designed for 15k images/month
PROJECT 04
Airbnb Translation
Enterprise real-time translation across 76 languages
A multilingual inference system designed around model-quality routing, large-scale caching, and a p90 real-time latency target.
TranslateGemma 27BRLHFvLLMB200RedisBedrockEKS
01Benchmarked models across 76 x 76 language pairs on eight B200 GPUs and selected TranslateGemma 27B with Claude Sonnet 4.5 fallback.
02Built a request path from a 1.5 TB Redis cache to accuracy-metric JSON routing, then to the best available translation model.
03Prepared RLHF data with Claude Opus 4.5, implemented reward-based training with model parallelism and vLLM forward passes, and deployed inference on EKS.
95% model accuracy
76 languages
1.5 TB Redis cache
1,500 ms p90 delivered
PROJECT 05
AI-Powered PCB Defect Inspection
Real-time manufacturing quality intelligence
Automated visual inspection for a factory producing 20,000 PCB boards each day across ten production lines.
Computer VisionSFTPydantic AIPostgreSQLVector SearchPython
01Built a ten-class defect detector using client and externally sourced data, expanded through data augmentation and supervised fine-tuning.
02Triggered a Pydantic AI agent after detection, passing the image and predicted class into a remediation workflow.
03Stored known-error knowledge in PostgreSQL with vector search so the agent can recommend line stops, machine routing, or corrective action.
96% defect accuracy
20k boards/day
10 production lines
10 defect classes
Technical toolkit
Breadth across the AI product stack.
AI & ML
Deep LearningMachine LearningNLPLLMsComputer VisionRAGSFTRLHFPrompt Engineering
Agentic AI
StrandsPydantic AILangChainLangGraphLangSmithAWS AgentCore
Backend & Data
PythonFastAPISQLNoSQLPostgreSQLMongoDBMySQLVector DatabasesWebSocket
Cloud & MLOps
AWS LambdaS3SQSECSEKSEC2BedrockSageMaker AIKubeflowDatabricksDockerKubernetesCI/CD
Frontend & Quality
ReactTypeScriptJavaScriptHTMLCSSPostmanThunder ClientTesting
Credentials
Education & certification.
2021 - 2025 | Chennai, India
B.E. Computer Science
Saveetha Engineering College
Professional certification
Databricks Certified Machine Learning Professional
Professional-level machine learning engineering credential.
Contact
Let's build something intelligent.
Reach out for enterprise AI engagements, consulting, or collaboration.
Direct
jagansivakumaran@gmail.comResponse Time< 4 hours
LocationChennai, India
SpecializationEnterprise AI systems