AI & Machine Learning Engineer

Bernard Birendra Das

Deep Learning LLM Pipelines Distributed MLOps

View Resume Explore AI Systems

WORK EXPERIENCE

2023 – 2025
Simplify3x Software Private LimitedSoftware Engineer — Bengaluru, Karnataka

ACADEMIC EXPERIENCE

2024 – Present
CHRIST (Deemed-to-be-University), Bangalore, KarnatakaPhD in Data Science
2021 – 2023
CHRIST (Deemed-to-be-University), Bangalore, KarnatakaMCA with Data Analytics
2018 – 2021
St. Xavier's College (Autonomous), Kolkata, West BengalBSc (Honours) Computer Science with Physics and Mathematics
2004 – 2018
Auxilium Convent School, Barasat, West BengalHigh School (ISC & ICSE)

About

What I do?

I am an Indian Citizen who is also a Software Engineer with 2+ years of enterprise experience.

I enjoy working on problems at the intersection of AI research and production systems. I specialize in high-throughput backend services and machine learning infrastructure.

Federated LearningFoundation ModelsHealthcare AIComputer VisionSignal ProcessingRegional Language Models

Work

Featured Systems

6 ARCHITECTURES//PRODUCTION GRADE
SYS_01AGENTIC AI & KNOWLEDGE GRAPHS
<120ms TTFT // 94.2% Multi-Hop Precision

Autonomous Multi-Agent GraphRAG & Financial Intelligence Engine

Production-grade hierarchical multi-agent network orchestrated with DSPy and LangGraph. Ingests raw SEC 10-K/10-Q filings, earnings transcripts, and live market news feeds. Features self-correcting query decomposition, ColBERT hybrid dense-sparse retrieval, knowledge graph traversal in Neo4j, and structured JSON output verification.

  • Multi-hop reasoning with GraphRAG resolving complex cross-document financial dependencies
  • Hybrid dense-sparse reranking combining ColBERTv2 with SPLADE for high recall
  • Dynamic agent tooling with automatic retry and schema-constrained decoding
LangGraphDSPyColBERTv2Neo4jQdrantvLLMSGLangDocker
GITHUB REPO
SYS_02DISTRIBUTED HPC & POST-TRAINING
48.2% MFU // 32x H100 Scale // 32k Long-Context

Distributed 3D-Parallel Foundation Model Pre-Training & Alignment Cluster

Scalable distributed training harness engineered for pre-training and alignment of a 14B Mixture-of-Experts (MoE) foundation model. Coordinates 3D parallelism across multi-node GPU clusters using PyTorch FSDP, DeepSpeed ZeRO-3, and Megatron-LM. Fine-tuned with Direct Preference Optimization (DPO) and online PPO/GRPO reinforcement learning.

  • Achieved 48.2% Model FLOPs Utilization (MFU) with zero gradient divergence across 32k context lengths
  • Automated activation checkpointing and ZeRO-Offload eliminating OOM bottlenecks
  • Integrated synthetic data generation and automated safety/bias evaluation benchmarks
PyTorch FSDPDeepSpeed ZeRO-3Megatron-LMFlashAttention-3Ray TrainTRLSlurmNCCL
GITHUB REPO
SYS_03INFERENCE ACCEL & CUSTOM KERNELS
4.2x Throughput // 68% p99 Latency Drop // 38% Cache Hit

Ultra-Low-Latency Real-Time LLM Serving Mesh with Speculative Decoding

High-throughput distributed model serving pipeline built for real-time generative workloads. Combines vLLM continuous batching and chunked prefill with speculative draft decoding (1.5B draft assisting 70B target). Features custom OpenAI Triton kernels for fused rotary embeddings and dynamic FP8/AWQ quantization.

  • 4.2x throughput increase over baseline serving with sub-32ms per-token generation latency
  • Custom OpenAI Triton kernels fusing layer norms and dynamic quantization for memory bandwidth savings
  • Redis-backed semantic response cache reducing redundant GPU compute by 38%
vLLMTensorRT-LLMOpenAI TritonCUDA C++RedisPrometheusGrafanaC++20
GITHUB REPO
SYS_04CLOUD MLOPS & PRODUCTION INFRA
99.99% Uptime // Sub-Second GPU HPA // Zero-Downtime Rollouts

Self-Healing Enterprise MLOps Platform & Continuous Evaluation Mesh

Cloud-native automated MLOps platform managing the entire model lifecycle from automated data validation to zero-downtime canary deployments on Kubernetes. Features real-time feature and concept drift monitoring that automatically triggers active learning re-training pipelines, backed by automated LLM-as-a-Judge evaluation suites.

  • Automated drift detection via Evidently AI triggering zero-touch re-training in Argo Workflows
  • Sub-second Horizontal Pod Autoscaling (HPA) using custom GPU utilization metrics on KEDA
  • Automated rollback mechanisms preventing production degradation during model upgrades
KubernetesKServeKubeRayMLflowArgo WorkflowsEvidently AILangfuseAWS SageMaker
GITHUB REPO
SYS_05EDGE AI & EMBEDDED ACCELERATION
60 FPS Real-Time (<18ms) // 75% Memory Cut // <25W Power

Multi-Modal Vision-Language Edge Perception System

Ultra-efficient edge vision-language reasoning engine deployed on NVIDIA Jetson AGX Orin for autonomous robotics and situational awareness. Features 4-bit AWQ weight quantization, TensorRT acceleration, open-vocabulary object grounding, and temporal video stream comprehension under strict embedded power constraints.

  • Sub-18ms end-to-end perception latency at 60 FPS running entirely within a 25W power envelope
  • 75% memory footprint reduction via 4-bit INT4 quantization without loss in mean Average Precision (mAP)
  • Direct integration with ROS2 middleware for real-time sensor fusion and spatial tracking
PyTorchTensorRTAWQOpenCVONNX RuntimeFastSAMROS2Modern C++
GITHUB REPO
SYS_06STREAMING DATA & HIGH-THROUGHPUT RECSYS
<18ms p99 at 75k QPS // Feature Freshness <150ms

Real-Time Streaming Feature Store & Dual-Stage Recommendation Mesh

High-velocity personalization pipeline processing 100k+ events/sec. Dual-stage retrieval architecture: candidate generation via Two-Tower embedding indexing in ScaNN, followed by deep transformer ranking (DLRM) served through Triton Inference Server, with strict point-in-time correctness to eliminate train-serve skew.

  • Processes 75,000 queries per second with p99 latency strictly below 18ms
  • Sub-150ms real-time feature updates using Apache Kafka and Apache Flink stateful streams
  • Feast-managed feature registry guaranteeing zero data leakage between training and serving
Apache KafkaApache FlinkFeastRedisTriton Inference ServerPyTorch DLRMScaNNDuckDB
GITHUB REPO

Skills

My Arsenal

8 DOMAINS//70+ PRODUCTION SKILLS
01CORE DL

Deep Learning & Neural Foundations

PyTorchJAX / FlaxFlashAttention-2/3Transformer InternalsTensorFlow / Keras 3Mixture of Experts (MoE)State-Space Models (Mamba)Autograd & Gradient DynamicsContrastive & Self-Supervised LearningReinforcement Learning (PPO, DPO, GRPO)NumPy & SciPy
02GEN AI

Generative AI, LLMs & Alignment

LLM Pre-training & SFTPEFT (LoRA, QLoRA, DoRA)Direct Preference Optimization (DPO)Hugging Face (Transformers, PEFT, TRL)Vision-Language Models (VLMs / Multimodal)Diffusion Models & Stable DiffusionRLHF & Reward ModelingTokenization & Vocabulary CurationModel Merging (MergeKit)
03AGENTS & RAG

Agentic Frameworks, Reasoning & RAG

LangGraphLlamaIndexDSPyGraphRAG (Neo4j)AutoGen & Multi-Agent TeamsHybrid Dense-Sparse RAG (ColBERT, SPLADE, BM25)Cross-Encoder Contextual RerankingStructured Outputs (Outlines, Instructor, SGLang)Self-Correcting Query DecompositionTool Use & Function Calling
04DISTRIBUTED HPC

Distributed Training & High-Performance Compute

PyTorch FSDPDeepSpeed (ZeRO-1/2/3)Megatron-LM3D ParallelismTensor & Pipeline Parallelism (TP/PP/SP)Ray Train & Ray CoreNCCL & InfiniBand TuningSlurm Cluster SchedulingGPUDirect RDMA & RoCEActivation Checkpointing & Memory Profiling
05SERVING & ACCEL

Inference Acceleration & Custom Kernels

vLLM (PagedAttention)TensorRT-LLMOpenAI Triton KernelsSpeculative DecodingTriton Inference ServerSGLang Runtime EngineQuantization (AWQ, GPTQ, FP8, INT4)Continuous Batching & Chunked PrefillTorchInductor / TorchDynamo (PyTorch 2.x)ONNX Runtime & TensorRT
06MLOPS & CLOUD

Production MLOps, Orchestration & Cloud

Kubernetes (K8s, KServe, KubeRay)AWS (SageMaker, Trn1/Inf2/p5)GCP (Vertex AI, Cloud TPU)Docker & Containerized MicroservicesMLflow & Weights & Biases (W&B)Argo Workflows & AirflowDrift & Data Quality Monitoring (Evidently AI)LLM Observability (Langfuse, TruLens)Prometheus, Grafana & OpenTelemetry
07DATA & STORAGE

Vector Search, Streaming & Data Infrastructure

QdrantMilvuspgvectorApache KafkaApache SparkPinecone & WeaviateApache Flink & Real-Time StreamsRay Data & Distributed IngestionDuckDB & Embedded AnalyticsApache Iceberg & Delta LakeRedis (Semantic Caching & Feature Buffers)
08LANGUAGES & SYS

Languages, Systems & Software Engineering

Python (Async, CPython Internals)Modern C++ (C++17/20)CUDA C/C++RustSQL (Advanced Analytical & Window Functions)Linux Kernel & POSIX Systems InternalsBash & Automated Infrastructure ScriptingGit, GitHub Actions & Enterprise CI/CDTypeScript & Modern Systems APIsHigh-Performance Multi-Threading

Contact

GitHubgithub.com/bernardbdas
LinkedInlinkedin.com/in/bernardbdas
Emailbernardbdas@gmail.com