Skip to content
View vinzlercodes's full-sized avatar
⛰️
Scaling new Goals
⛰️
Scaling new Goals

Block or report vinzlercodes

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
vinzlercodes/README.md

Hi, I'm Vinayak Sengupta

AI-Native Product & Platform Engineer | LLM Systems | Data Science

I build production AI systems that turn model capability into usable enterprise workflows: agent orchestration, fine-tuning, retrieval evaluation, explainability, analytics products, and the platform plumbing needed to keep all of it observable.

LinkedIn Medium GitHub Email Resume


What I Build

Area Current focus
Agent platforms Multi-agent coordination, deterministic tool execution, human-in-the-loop confirmations, tool-call graphs, and enterprise workflow automation.
LLM operations SFT, DPO, PPO, QLoRA, Axolotl, vLLM, checkpoint recovery, model serving, and failure diagnostics for production AI workflows.
Retrieval and evaluation RAG/GraphRAG over long-form and graph-structured documents, ranking quality, BM25, MMR, reranking, and nDCG@k measurement.
Explainability and analytics Model-agnostic attribution, PDP workflows, KPI design, DuckDB/Arrow pipelines, SHAP alignment, and stakeholder-facing analytics products.
Platform engineering FastAPI services, Kubernetes reconciliation loops, productized launch flows, health checks, observability, and reusable enterprise assets.

I care about systems that are measurable, debuggable, and useful to the people who have to run them after the demo is over.

Recent Work

Aible - Data Scientist, AI-Native Systems & Enterprise Workflows

  • Architected an NVIDIA NeMo Agent Toolkit LLM runtime for dynamic reasoning chains, multi-agent coordination, deterministic tool execution, and enterprise workflow automation.
  • Built a FastAPI execution layer that converts user/model intent into runnable agent configurations across tools, document workflows, prediction scoring, and human-in-the-loop approvals.
  • Designed a pluggable enterprise tool registry with persisted execution metadata and tool-call graphs for observability, reproducibility, workflow reuse, and operational debugging.
  • Productized an OpenClaw-based agent platform with sandbox launch flows, tenant configuration, health checks, gateway reachability checks, Slack access, streaming responses, and failure diagnostics.

AI Operations, Fine-Tuning & Reliability

  • Led a fault-tolerant fine-tune-and-serve platform for enterprise AI use cases, translating customer requirements into scalable experimentation and deployment workflows.
  • Implemented SFT and DPO workflows using Axolotl and QLoRA; automated checkpoint detection and recovery, reducing manual setup and monitoring effort by 80%.
  • Enabled a Fortune 50 telecom client to launch a security metadata classifier on schedule through a productionized fine-tuning and serving workflow.

Retrieval, Document Intelligence & Explainability

  • Designed retrieval workflows for graph-structured and long-form enterprise documents while balancing ranking quality, token constraints, modular experimentation, and production evaluation.
  • Improved retrieval nDCG@k by 25% through iterative tuning of BM25, Maximal Marginal Relevance, and reranking components.
  • Led a standardized model explainability/PDP analytics workflow that reduced per-feature computation time by 17x while maintaining median curve fidelity around 0.90.

Current Product Builds

Build What it explores
Open Prior Auth Workbench A FHIR-first healthcare AI workbench for discovering coverage requirements, prefilling documentation questionnaires, assembling submission-ready packets, and tracking case status through human review.
Multi-agent commerce systems A Swiggy-style production multi-agent system spanning food, delivery, and dine-out domains.
LLM workflow tools Obsidian/n8n LLM-wiki writing agents, MiroFish-style LLM councils for risk decisioning, GEN-1 robotics concepts, and Codex agentic OS experiments.

Technical Toolkit

Agentic AI & LLM systems multi-agent orchestration · tool registries · AgentOps · human-in-the-loop flows · RAG · GraphRAG · SFT · DPO · PPO · QLoRA · Axolotl · vLLM · LangChain · LlamaIndex · NVIDIA NeMo Toolkit · NeMo Guardrails · OpenAI · Vertex AI

Platforms, data & backend Python · SQL · Cypher · FastAPI · Flask · PySpark · DuckDB · PostgreSQL · MongoDB · Neo4j · Chroma · AWS · GCP · Docker · Kubernetes · GitHub Actions · OpenTelemetry · Langfuse

ML, product & analytics PyTorch · TensorFlow · Keras · scikit-learn · LightGBM · SHAP · ONNX · PDP · KPI design · stakeholder discovery · PRDs · MVP roadmaps · success metrics

Selected Public Projects

Project Signal
Gaming-Industry-Analysis Data analysis of a 40-year gaming dataset, including genre/platform trends, sales patterns, publisher contributions, and a companion long-form article.
Prediction-of-Customer-Churn ANN-based churn prediction for banking customers with ROC, confusion matrix, pie chart, KDE, and counter-plot analysis.
Disaster-Response-Pipeline-Web-App End-to-end ETL, NLP, and ML pipeline powering a web app for classifying disaster-response messages.
Recommendation-of-Refactoring-Techniques-to-address-Self-Admitted-Technical-Debt SATD detection and refactoring recommendation work from my RIT capstone.

Writing & Research

Auto-updated from my Medium RSS feed.

Other work: PPO post-training for Llama text-to-SQL, SATD detection and refactoring recommendation, and histopathology carcinoma classification using multi-level spatial fusion.

Talks & Community

  • Authored the core problem statement and evaluation metrics for the UC Berkeley AI Summit 2023 - Data Science Hackathon.
  • Represented Aible at Ai4 2023, Google Next 2024, and AWS Summit 2024, translating technical systems into demos and customer conversations.
  • Write long-form pieces on document summarization, retrieval systems, RAG evaluation, customer segmentation, applied AI, and gaming industry analysis.

GitHub Activity & Analytics

Generated GitHub metrics for vinzlercodes, including contribution activity, repository counts, community stats, and most-used languages

Signal What to look for
Languages A practical mix of data, backend, notebooks, and web-facing work rather than a single narrow stack.
Repositories Public projects skew older but show the arc from analytics and ML pipelines toward AI-native systems.
Writing Medium activity makes the technical reasoning visible, especially around retrieval, summarization, and applied analytics.
Activity feed Recent public GitHub events are generated below so profile movement is visible between larger project updates.

Recent GitHub Activity

  1. 🎉 Merged PR #45 in vinzlercodes/DecisionRisk
  2. 🎉 Merged PR #46 in vinzlercodes/DecisionRisk
  3. 🎉 Merged PR #47 in vinzlercodes/DecisionRisk
  4. 🔒 Closed issue #10 in vinzlercodes/DecisionRisk
  5. 🗣 Commented on #10 in vinzlercodes/DecisionRisk
  6. 💪 Opened PR #47 in vinzlercodes/DecisionRisk
  7. 💪 Opened PR #46 in vinzlercodes/DecisionRisk
  8. 💪 Opened PR #45 in vinzlercodes/DecisionRisk
  9. 🎉 Merged PR #44 in vinzlercodes/DecisionRisk
  10. 🎉 Merged PR #43 in vinzlercodes/DecisionRisk

Fun fact: I will absolutely over-analyze both fragrance notes and video-game industry trends.

Pinned Loading

  1. DecisionRisk DecisionRisk Public

    DecisionRisk: simulate, debate, and quantify the downside of consequential decisions before acting.

    Python

  2. Open_Prior_Auth_Workbench Open_Prior_Auth_Workbench Public

    A FHIR-first prior authorization workbench that discovers coverage requirements, retrieves and prefills documentation questionnaires, assembles submission-ready packets, and tracks lifecycle state …

    TypeScript