Agentic AI & LLMs
Autonomous agents, multi-agent orchestration, RAG, and fine-tuning.
LangGraphLangChainLlamaIndexvLLMLoRA / QLoRAQuantization
Samir Sengupta
Open to work
Samir Sengupta - AI/ML Engineer & Data Scientist in New York, USA
I frame the problem, build the model, and ship the system.
Multi-agent systems, RAG, and LLM pipelines, engineered onto Kubernetes with the data and evaluation infrastructure to keep them honest.
- 500K+
- <100
- ~40%
- ~32%
stack
- Python
- PyTorch
- LangGraph
- vLLM
- Kubernetes
- AWS Bedrock
- Apache Beam
- Kafka
- FastAPI
- Azure OpenAI
- FAISS
- MLflow
intermission · the input device
press any key
Every pipeline on this page started at a keyboard. This one works: scroll lifts the keys off the plate, and typing plays it, sound on.
One engineer, the full ML lifecycle.
portrait.jpg
Most engineers pick a lane - research, modeling, or infrastructure. I work the whole line. Across 3+ years I’ve framed problems as a data scientist, built the models, and engineered them into reliable software: Kafka pipelines and FastAPI services on Kubernetes at sub-100ms latency, and predictive models that cut churn and lifted conversion for the teams that shipped them.
I’m the author of a research preprint on efficient long-context LLMs, spoke at KCD New York and the Apache Beam Summit in 2026, and hold an MS in Data Science (GPA 3.9) plus industry certifications from AWS, Google, NVIDIA, and IBM.
framemodelship
ML & Data Science
Forecasting, anomaly detection, churn and predictive modeling, A/B testing.
scikit-learnXGBoostPyTorchTensorFlowPython & MLOps
FastAPI services, event-driven microservices, CI/CD, and observability.
KubernetesDockerMLflowAWS BedrockAzure OpenAI
Where I’ve shipped.
Software DeveloperSynRadar
- Deployed FastAPI backend services on Kubernetes via CI/CD pipelines, serving ML predictions at sub-100ms latency while handling ~50K daily API calls.
- Engineered a Kafka-based data pipeline integrating external security APIs - managing rate limits, retries, and normalisation to process ~500K daily events.
- Integrated a Llama-based LLM into the automated review workflow, with structured logging and tracing to monitor model decisions and detect regressions.
Data ScientistNeural Thread
- Built a Python NLP pipeline classifying support tickets and integrating with business systems to automate routing, cutting manual triage time ~40%.
- Trained churn-prediction models on SQL-extracted usage data, helping retention teams flag at-risk accounts and reduce monthly attrition ~25%.
- Ran A/B testing infrastructure for targeted campaigns, measuring predictive-model impact to drive a ~32% improvement in conversion rates.
education
M.S. Data Science
Saint Peter’s University
New JerseySep 2024 - May 2026GPA 3.9 / 4.0B.S. Data Science
University of Mumbai
Mumbai2020 - 2023GPA 3.95 / 4.0
How a model becomes a system.
A production LLM/RAG pipeline I design and operate - from raw data to a served, monitored endpoint that stays fast and cheap under load. Pick a stage.
01 / 07
Ingest & stream
metric500K events/day
Event-driven intake that survives the real world: external APIs behind rate limits, retries with backoff, schema normalisation. ~500K events a day arrive without anyone getting paged.
- Kafka
- Apache Beam
- Airflow
Published, in the open.
Preprint · TechRxivJanuary 2026
HLST: Democratizing Long-Context LLMs on Edge Devices via Hierarchical Latent State-Space Recurrence
A transformer alternative built on hierarchical latent state-space recurrence - trading quadratic attention for recurrence that holds long context on hardware that fits in a pocket.
- Long-context
- State-space models
- Edge inference
- Efficiency
doi10.36227/techrxiv.176948384.49791660/v1
Presented this line of work on stage in 2026 - see below.
The toolkit.
Grouped the way the system is layered, not alphabetically.
- LangChain
- LangGraph
- LlamaIndex
- CrewAI
- RAG
- Fine-Tuning
- LoRA / QLoRA
- vLLM
- DSPy
- Guardrails.ai
- PyTorch
- TensorFlow
- scikit-learn
- XGBoost
- Pandas
- Forecasting
- Anomaly Detection
- A/B Testing
- NLP
- Computer Vision
- Kubernetes
- Docker
- Terraform
- MLflow
- Ray Serve
- Ollama
- CI/CD
- Model Monitoring
- Apache Spark
- Apache Beam
- Airflow
- Snowflake
- AWS Bedrock
- SageMaker
- Azure OpenAI
- GCP Vertex AI
- FAISS
- Pinecone
- ChromaDB
- Elasticsearch
- Hybrid Search
- Embeddings
- Python
- SQL
- C++
- Java
- TypeScript
- Rust
- Go
10+ certifications.
AWSGoogleMicrosoftNVIDIATensorFlowIBMPyTorchStanford
- AWSAWS ML - Specialty
- GoogleProfessional ML Engineer
- MicrosoftAzure AI Engineer
- NVIDIAAgentic AI
- NVIDIAGenAI & LLMs
- TensorFlowTensorFlow Developer
- IBMAI Engineering
- PyTorchDeep Learning
- StanfordAI Graduate Program
- IBMRAG & Agentic AI
Projects, shipped on GitHub.
Ten of them, all public - nine on GitHub and one on the VS Code Marketplace.
1010 projects shownselected work
- featured · newWayneView project01 of 10
Wayne
AI harness for VS Code
An AI harness living in the IDE - it plans before it codes, reads the codebase and follows call sites, runs commands and reacts to failures, and presents every change as a diff you approve. Extends through MCP. Live on the VS Code Marketplace.
- flagshipPrometheusAIView project02 of 10
PrometheusAI
Offline LLM assistant for Android
On-device inference with zero cloud dependency. Runs LLaMA 3.1, IBM Granite, and Mistral via 4-bit GGUF at ~8 tok/s, with on-device RAG over your own documents.
- recenttokenashView project03 of 10
tokenash
Local-first context compression
Context compression for LLM applications that runs locally - 50–90% fewer tokens for the same answers, with nothing shipped to a third-party service to get there.
- preprintHLSTView project04 of 10
HLST
Long-context LLMs on edge
Reference code for my TechRxiv preprint: hierarchical latent state-space recurrence that democratises long-context LLMs on edge devices, as a transformer alternative.
- talkKCD New York 2026View project05 of 10
KCD New York 2026
Production RAG on Kubernetes
Demo from my KCD New York 2026 talk - deploying, autoscaling, and observing production RAG pipelines on Kubernetes with vLLM and vector search.
- talkApache Beam Summit 2026View project06 of 10
Apache Beam Summit 2026
Real-time AI pipelines
Embedding LLMs and RAG directly into Apache Beam transforms for low-latency inference on high-velocity data streams.
- RelayView project07 of 10
Relay
Open-source outbound agent
An open-source AI agent for outbound leads - the research, qualification, and follow-up loop handled by a tool-using agent rather than a sequence of manual steps.
- FranolaView project08 of 10
Franola
Open-source AI note-taking
An open-source AI note-taking app in the Granola mould, written in Rust - meeting notes captured and structured by a model, without the notes leaving a proprietary product.
- GateKeepView project09 of 10
GateKeep
Agentic personal finance
An agentic-AI finance tracker that ingests transactions and reasons over them with tool-using agents to surface spend insights.
- AIORNORView project10 of 10
AIORNOR
AI-or-human text detection
Detects whether a piece of text is AI-generated or human-written using a perplexity-based metric, served through a lightweight web app.
Measured, not vibes.
Numbers from systems that shipped and stayed up. Every one of them came with a dashboard someone else had to trust.
- ~95%Higher is better.
enrichment success
multi-agent + retries
- 500K+Higher is better.
events / day
Kafka pipeline
- <100msLower is better.
prediction latency
FastAPI on Kubernetes
- ~40%Lower is better.
manual triage cut
NLP ticket routing
- ~25%Lower is better.
monthly churn reduction
predictive ML
- ~32%Higher is better.
conversion lift
A/B testing programme
Taking production AI to the stage.
In 2026 I took the hard part of applied AI to two New York stages - not the demo, but making it run reliably, cheaply, and at scale.
- 06.10.262026Lightning talkrecording live
Scaling Production RAG Systems with Kubernetes
How retrieval-augmented generation earns its uptime: autoscaling vLLM, vector search, and observability on Kubernetes so a RAG system stays fast and affordable when traffic stops being a benchmark and starts being real.
- 06.22.262026Technical sessionrecording live
Real-Time AI Pipelines at Scale: Embedding LLMs into Apache Beam
Bringing LLM inference and RAG inside the pipeline itself - embedding models directly into Apache Beam transforms so streaming data can reason in real time, delivering low-latency intelligence on high-velocity data without ever leaving the stream.
- 09.16.262026Conference sessionupcoming · Sep 16–17, 2026
Session To Be Announced
I will be speaking at Blockchain Week - UNGA Edition 2026, the ten-day independent industry gathering held in New York alongside the UN General Assembly (September 10–19), with tracks spanning Bitcoin, AI agents, energy, and the space economy. The conference will announce the session title and slot; the talk sits where my work does - production AI systems, and what it takes to run them reliably once the demo is over.
I write, and it gets read.
- Gemini 3.8 Flash Cyber benchmark performance on vulnerability detection4 min readSamCodeMan.
- Run 104GB Qwen3.8-Flash-Next on a 48GB Mac via slotstream at ~12 tok/s8 min readSamCodeMan.
- 44% on ARC-AGI-1 for 67 cents with a test-time-trained transformer5 min readSamCodeMan.
- Claude Code Auto Mode refused the binary and then ran its own poisoned decoder5 min readSamCodeMan.
- vLLM v0.28.0 doubles the default batched-token budget and drops bitsandbytes in-tree4 min readSamCodeMan.
- musl with mimalloc is still 26% slower than glibc on a 4-core box4 min readSamCodeMan.
- GLM-5.3 ships open weights with no published parameter count5 min readSamCodeMan.
daily notes · on this site
People vouch for the work.
A mentor, a Microsoft engineer, and a reader, in their own words.
Samir is an inventor. His work with agentic AI, open-source contributions on Hugging Face, and a locally-run IBM Granite LLM app for Android set him apart. There were times you’d be hard-pressed to tell who was the professor.
Armen PischdotchianMentor · Adjunct Professor, Saint Peter’s University Our team had spent weeks trying to structure a complex agentic-AI workflow. Samir stepped in, grasped the entire system, and had it fully running in under a week.
Mitesh WaghelaSupport Escalation Engineer, Microsoft LOVED your productivity paradox feature. Just read your Medium article and I loved it - particularly the guilt piece! I talk a lot about maximising our human capacity to thrive in this era, and that was a new one.
Suzanne RathKeynote Speaker · High Performance Consultant
Let’s ship AI that survives production.
Open to AI/ML Engineering, Data Science, and Python roles, plus research collaborations and consulting. Based in New York, shipping worldwide.
Fastest route - straight to my inbox