Trayan Das

AI/ML Computational Science Specialist — Accenture

Sep 2026 – Present

• Forensic Knowledge Assistant (Multimodal RAG Agent): Designed and built a multimodal, citation-grounded Q&A assistant (Microsoft Foundry, Azure AI Search, Azure Functions, Copilot Studio, Power Platform) answering engineering-investigation questions from text, an image, or both in Teams and Copilot chat. Enforced grounding in code under a cite-or-silent policy — every cited chunk ID validated against the retrieved set, with a deterministic "not found" fallback — and secured it with Microsoft Entra ID (OAuth 2.0) and a least-privilege managed identity. Stage-wise evaluation: recall@10 0.84, RAGAS factual correctness 0.87, 2% false-answer / 5% false-refusal, p95 8s text-only / 20s with image — judged by a model from a different family than the one under test.

Software Developer — Expro

Sep 2022 – Sep 2024

• Well Performance Analytics Platform: Implemented a PyTorch LSTM-autoencoder pipeline for anomaly detection on well-sensor time-series data, improving risk detection by ~20% and reducing analysis time by ~40%.
• Oil & Gas Insights Web Application: Designed a PySpark KPI pipeline on Azure Databricks using PCA, K-means, and Random Forest to derive KPI features from equipment telemetry, reducing manual reporting by ~50%.

Data Scientist — Ganit Labs

Mar 2021 – Sep 2022

• Developed predictive maintenance models on GCP for oilfield equipment, achieving ~85% recall and reducing downtime by ~30%.

RTI-Engine — AI Pay Transparency Compliance Engine

An 11-node LangGraph multi-agent system for pay-transparency compliance — 6 model-driven nodes and 5 deliberately deterministic ones where no judgment is required. Tiered autonomy enforced in code, conditional human approval gates, cross-vendor model review, and deterministic figure-validation. Data exposed through MCP servers over a Neo4j knowledge graph parsed from directive text; deployed on Azure with Docker and GitHub Actions CI/CD, instrumented with LangSmith and OpenTelemetry tracing plus an openevals harness.

View code → Read the write-up →

Autonomous Data Analyst

A LangGraph multi-agent text-to-SQL system on Azure with governance-first routing, execution-repair SQL validation, forced root-cause decomposition, model tiering for cost control, and OpenTelemetry cost tracing — 87% accuracy on a 110-question golden-set eval harness.

View code → Read the write-up →

MARA — Multimodal Agentic Reasoning Assistant

A multi-agent system coordinating RAG, Vision, Data, and Web Search agents for multi-modal queries, fanning out 3–5 agents in parallel per query. Hybrid FAISS + BM25 retrieval at recall@10 ~0.75, adaptive planning with error recovery, served over FastAPI at p95 ~30s end to end.

View code → Read the write-up →

GPT-Style Language Model from Scratch

A decoder-only 124M-parameter LLM implemented from scratch in PyTorch — BPE tokenization, multi-head causal self-attention, the training loop, and instruction fine-tuning — trained on 124M tokens.

View code → Read the write-up →

Sovereign-Policy-Assistant

Sovereign, offline RAG policy assistant with LangGraph, bilingual retrieval, Keycloak-gated fully offline confidentiality, and contradiction/staleness detection.

View code →

Hybrid RAG from Scratch

Hybrid retrieval built from first principles — from-scratch BM25, reciprocal rank fusion, FAISS vector search, and RAGAS evaluation.

View code →

Smart Compose System

Fine-tuned GPT-style model powering intelligent email composition.

View code →

LLM Summarization Pipeline

End-to-end summarization workflow using modern LLM techniques.

View code →

Company Brochure Generator

Automated content generation pipeline using large language models.

View code →

Languages: Python, SQL

AI Architecture & Governance: Solution architecture, multi-agent & RAG design patterns, AI guardrails, human-in-the-loop governance, cost/latency trade-off analysis

Generative AI: Multi-agent orchestration, Retrieval-Augmented Generation (RAG), prompt engineering, LLM fine-tuning (LoRA), AI observability, AI evaluation, Model Context Protocol (MCP)

Frameworks & Libraries: LangChain, LangGraph, CrewAI, Claude Agent SDK, Azure OpenAI / OpenAI SDK, Hugging Face Transformers, PyTorch, scikit-learn, FastAPI, Flask

Machine Learning & Deep Learning: LSTM autoencoders, Isolation Forest, PCA, K-means, HDBSCAN, Random Forest, Decision Tree, Linear & Logistic Regression, NLP

MLOps & DevOps: Docker, Git, Claude Code, GitHub Actions CI/CD, Azure DevOps, Bicep IaC, pylint, MLOps & LLMOps pipelines, LangSmith, OpenTelemetry, Application Insights (KQL), OpenEvals, RAGAS, REST APIs

Cloud Platforms: Azure (AI Foundry, Storage, Fabric, Cosmos DB, AI Search, OpenAI, App Service, Function App, Logic Apps, Container Apps); GCP (Cloud Storage, BigQuery, App Engine)

Data & Retrieval: Vector search — FAISS, Pinecone, Azure AI Search; Graph — Neo4j; Platforms — Microsoft Fabric, Azure SQL, Azure Databricks

• Spotlight Award — Client, Feb 2026, for delivery of the privacy-safe analytics agent platform
• Growth Catalyst Award — Accenture, Dec 2025, for technical proficiency and project delivery
• Promoted within 10 months at Ganit Labs
• Published technical write-ups on all four flagship projects — see medium.com/@trayandas

AI-102: Microsoft Azure AI Engineer Associate

AZ-900: Microsoft Azure Fundamentals

AI-900: Microsoft Azure AI Fundamentals

Claude Certified Architect – Foundations

CS50AI: Introduction to Artificial Intelligence with Python — Harvard University

LLMOps, MLOps & CI/CD/CT Pipelines — Udemy

M.Tech — University of Calcutta (2016–2018)

B.Tech — WBUT (2011–2015)