Senior Staff AI Engineer
Jayesh Bankoti
8 years building production AI — from narrative intelligence at scale to agentic cloud migration systems.
About
8 years building AI that ships.
I'm a Senior Staff AI/ML Engineer with 8 years of experience — 4+ at Staff level — turning research into production systems. I studied Mechanical Engineering at IIT Delhi and was a KVPY Fellow, but code was always the real obsession.
My career spans the full arc of modern ML: starting as a Data Scientist at VMock building knowledge graphs over 2,000+ skills, through NLP team leadership at Brane Enterprises, to building and running the entire AI/DS organization at Logically AI post-Series A — directly reporting to the CEO.
At Logically AI I drove 120% platform engagement growth, cut infrastructure costs by 82%, and built the narrative intelligence system that anchored a $24M Series A (Vitruvian Partners, Amazon Alexa Fund). Now at Tiger Analytics, I'm building LLM-powered cloud migration systems: AST-based code analysis, hybrid RAG chatbots, and autonomous code migration agents.
Skills
I make AI work — fast, smart, and scalable.
Production-hardened expertise across the full ML lifecycle.
GenAI Systems
- ▸LangGraph multi-agent pipelines with deterministic Python supervisors and adversarial debate loops.
- ▸Hybrid RAG: Table-Augmented Generation (DuckDB SQL), vector RAG, and LLM routers that pick dynamically.
- ▸Agentic chatbots with SmolAgents (ReAct), LangChain tools, and summarization-based multi-turn memory.
- ▸LLM fine-tuning: QLoRA + PEFT on GEMMA, two-stage SFT over 50K+ instruction pairs.
- ▸Code migration agents with generator-reviewer pattern and LLM-as-judge dry-run gating.
ML & NLP
- ▸Replaced E5 with Qwen3 LLM-based embeddings + GPU-accelerated UMAP/HDBSCAN + MinHash-LSH for CIB detection.
- ▸Built NER, sentiment, language detection, geolocation, and translation annotation pipelines at social-media scale.
- ▸BiLSTM-CNN resume classifier (F1 90%) and 2,000+ node skill knowledge graph with graph-algorithm ranking.
- ▸Treesitter AST extraction for code analysis across 9 cloud migration patterns (AWS, GCP, Azure, on-prem).
- ▸SentenceTransformers with FlashAttention and BF16 precision; 35% embedding latency reduction.
MLOps & Inference
- ▸45% inference latency reduction via ONNX Runtime + TensorRT with BF16 optimization.
- ▸PyTorch 2.x compiler stack (FlashAttention, BF16) for production SentenceTransformer inference.
- ▸Databricks CI/CD: Asset Bundles, service principals, strategic microservice-to-monolith consolidation.
- ▸Docker + FastAPI + gRPC production serving; MLflow experiment tracking.
- ▸Shifted orchestration from reserved to on-demand compute — 82% infrastructure cost reduction.
Data & Infrastructure
- ▸Hybrid data stack: PostgreSQL (OLTP), Delta Lake (OLAP), Elasticsearch (sub-second retrieval), MongoDB.
- ▸Neo4j knowledge graphs for cloud migration: service profiles, IaC mappings, database inventories.
- ▸DuckDB SQL over structured reports for Table-Augmented Generation (TAG) in hybrid chatbots.
- ▸Dataplex-generated rule validation with LLM-as-judge sign-off for post-migration verification.
- ▸PySpark + Databricks for large-scale data processing and analytics pipelines.
Experience
8 years, production systems, measurable impact.
- ▸Cloud migration assessment framework: Treesitter AST extraction feeding Neo4j graphs that auto-generate service profiles, IaC mappings, DB inventories, and change reports across 9 migration patterns (AWS, GCP, Azure, on-prem).
- ▸Hybrid chatbot with an LLM router that picks between Table-Augmented Generation (DuckDB SQL over structured migration reports), vector RAG, or both based on query type.
- ▸Code migration agent for AWS EMR-to-GCP: generator-reviewer pattern with an independent LLM-as-judge, dual-gated on dry-run execution before any change is accepted.
- ▸Post-migration rule validator where Dataplex-generated rules go through an independent LLM judge before sign-off.
- ▸Ran the 9-engineer AI/DS org as Senior Staff, the highest-scope technical role post-Series A. Owned roadmap, architecture, quarterly delivery, and reported directly to the CEO.
- ▸120%+ platform engagement growth: replaced E5 with Qwen3 LLM embeddings, moved to GPU-accelerated UMAP + HDBSCAN (Optuna-tuned), and added MinHash-LSH deduplication to surface coordinated inauthentic posting across millions of social posts.
- ▸82% infrastructure cost reduction: consolidated microservices into a single workflow and shifted from reserved to on-demand compute, with PostgreSQL handling state and messaging.
- ▸35% faster embedding generation via PyTorch 2.x compiler, FlashAttention, and BF16 precision.
- ▸Agentic narrative Q&A chatbot using SmolAgents (ReAct) with LangChain tools and Mosaic AI vector DB, with multi-turn summarization-based memory.
- ▸Migrated the AI platform to Databricks with Asset Bundle CI/CD and service principals, unifying DS, engineering, and MLOps into one delivery workflow.
- ▸Grew the Media Intelligence NLP team from scratch to 3 engineers. Owned hiring, mentoring, technical direction, and delivery.
- ▸Narrative Intelligence System that anchored the $24M Series A (Vitruvian Partners, Amazon Alexa Fund): narrative detection, annotation pipelines, and analytics in one platform.
- ▸45% inference latency reduction via ONNX Runtime + TensorRT with BF16 optimization, which became the production standard across the AI team.
- ▸End-to-end annotation pipelines for NER, sentiment, language detection, translation, and geolocation across all social post ingestion.
- ▸Led 12 NLP engineers across design, integration, and delivery of NLP/ML modules: information extraction, classification, and QA.
- ▸FAQ QA system using BM25 + RoBERTa fine-tuned on SQuAD 2.0, hitting 80% accuracy in internal eval and shipped into the production help-desk chatbot.
- ▸BiLSTM-CNN resume classifier (F1 90%) deployed across 4 live products used by 150+ universities including Stanford, NTU, and London Business School.
- ▸2,000+ node skill knowledge graph with graph-algorithm ranking for job recommendations at 90% relevancy, powering Wall of Jobs across 30+ colleges.
- ▸Job classification and recommendation models (CNN + collaborative filtering) for the core university-facing jobs platform.
- ▸Linear Regression model (85% accuracy) replicating the force-transmission behaviour of endoscopy devices.
- ▸MATLAB simulation with interactive GUI identified the primary cause of force loss, leading to a 35%+ improvement in force transmission.
Projects
Shipping ideas, on video.
End-to-end projects with walkthrough demos — not just README files.
Autonomous AI Research System
LangGraph multi-agent pipeline with a deterministic Python supervisor, adversarial two-agent debate loop (Research vs Critic), and isolated single-task SLMs for fault-tolerant research automation.
Hierarchical RAG for Financial Analysis
Hierarchical semantic RAG clustering 150–200 token fragments into ~1,500 token superchunks via HDBSCAN. FAISS HNSW retrieval with LLM query reformulation for citation-grounded generation.
Potterhead GPT
GEMMA fine-tuned with QLoRA via two-stage training: domain grounding on all 7 books, then SFT on 50K+ multi-persona Q&A pairs. Streaming inference with dynamic persona switching via FastAPI.
Research
Published & peer-reviewed.
Logically at the Constraint 2022: Multimodal Role Labelling
Constraint Workshop @ ACL 2022 · 2022
Ranked 2nd among 9 finalist systems from 100+ competing teams.
Certifications
Always learning.
Anthropic
Certificate of completion: Introduction to Subagents
Certificate of completion: Introduction to Agent Skills
Claude Code in Action
Certificate of completion: Claude Code 101
Educative
Machine Learning System Design
Data Science in Production: Building Scalable Model Pipelines
Applied ML in Python: NumPy, pandas & scikit-learn in Action
Applied ML for Engineers: Supervised Learning, Neural Networks & Deep Learning
Cracked AI Engineering Interviews: Advanced Strategies & Techniques
SQL for Data Professionals: Intermediate Querying & Analytics
Mastering Knowledge Graphs & RAG with Neo4j
DeepLearning.AI
Neural Networks and Deep Learning
LangChain for LLM Application Development
Natural Language Processing in TensorFlow
Building Your Own Database Agent
Functions, Tools and Agents with LangChain
Building Code Agents
TensorFlow for AI, ML, and DL
Hyperparameter Tuning, Regularization and Optimization
Convolutional Neural Networks
IISc
Deep Generative AI
Contact
Need AI help? Let's talk.
I'm open to senior/staff AI engineering roles and consulting on GenAI, NLP, or MLOps challenges. I reply within 24 hours.
jayeshbankoti@gmail.com
Phone / WhatsApp
+91 84483 59710
linkedin.com/in/jayeshbankoti
GitHub
github.com/jayeshbankoti007
Based in India (Remote-first)
