AI/ML ENGINEER · BENGALURU
I build practical AI systems across forecasting, agentic workflows, recommendation engines, and backend infrastructure.
Final-year CSE (AI/ML) student and ex-NetApp ML intern focused on production-oriented ML workflows, LangGraph agents, RAG evaluation, and scalable APIs.
Optional listen
A quick 45-second introduction to who I am and what I build.
Plays at 1.5xHi, I’m Shrey Singh, an AI/ML engineer based in India. I’ve worked on production ML systems at NetApp, including revenue forecasting pipelines, sales commission forecasting, and deal-scoring workflows. I like building practical AI products — things like AI waiters, RAG assistants, workflow automation agents, and tools that connect LLMs with real APIs. I’m especially interested in machine learning systems, GenAI applications, and agentic products that solve real business problems. You can explore my work below, and if something looks interesting, feel free to reach out.
A static-readable snapshot of target roles, internship impact, systems focus, and stack depth.
AI/ML Engineer, GenAI Engineer, MLOps Engineer, Backend Engineer
$150M+ quarterly planning, 1.5M+ records, 4,900+ sales reps, 15% accuracy lift, 6-10% MAPE
RAG, LangGraph agents, forecasting pipelines, CI/CD retraining, quant ML
Python, SQL, Go, TypeScript, FastAPI, LangGraph, XGBoost, Redis, Docker, CI/CD
A focused set of RAG evaluation, agentic automation, AI product, fairness planning, and MLOps systems. Public code is linked where verified; private resume work is marked honestly.
Recall@5 0.960, MRR 0.840, CI gate
Built a production-style RAG and agent evaluation lab with deterministic fixtures, retrieval metrics, citation-grounding checks, LangGraph tool workflows, and CI thresholds that fail on quality regressions.
Typed state, 75+ tests, audit logs
Built a LangGraph-based automation agent integrating Gmail and Google Calendar APIs to search email threads, summarize invoice context, draft reminder emails, and schedule follow-ups with human approval checkpoints.
65 backend tests, eval dashboard, citations
Built a production RAG assistant with PDF ingestion, semantic chunking, pgvector-based hybrid retrieval, citation-grounded answer generation, and FastAPI inference endpoints.
Function calling, menu cards, WhatsApp handoff
Built an AI restaurant ordering agent with Gemini function calling, FastAPI tool endpoints, structured menu schemas, allergy-aware recommendations, cart updates, repeat-customer personalization, and WhatsApp checkout handoff.
Fairness ledger, approval loops, dashboard
Built a LangGraph-based group planning engine with typed graph state, conditional routing, revision loops, persistent memory, and human approval checkpoints.
Drift checks, eval gates, rollback demo
Built an automated MLOps retraining pipeline with drift detection, evaluation gates, champion-challenger promotion, canary rollout, rollback handling, and GitHub Actions orchestration.
Filter by domain or stack. Every card here has a real system, proof point, or product angle.
Recall@5 0.960, MRR 0.840, CI gate
Customer-intelligence RAG evaluation lab with a deterministic 100-question benchmark, citation checks, LangGraph tool calls, and CI quality gates.
Typed state, 75+ tests, audit logs
LangGraph workflow agent that searches email threads, summarizes invoice context, drafts reminders, and schedules follow-ups with approval checkpoints.
65 backend tests, eval dashboard, citations
Production-grade RAG app for uploading PDFs, indexing documents, and asking citation-grounded questions through hybrid retrieval.
Function calling, menu cards, WhatsApp handoff
Conversational restaurant ordering MVP that recommends dishes, remembers preferences, updates carts, and hands off orders through WhatsApp.
Fairness ledger, approval loops, dashboard
Group planning engine that recommends fair meetups across location, interests, budget, energy level, and long-term compromise history.
Drift checks, eval gates, rollback demo
Production-style MLOps reference for drift detection, automated retraining, evaluation gates, and shadow/canary/blue-green deployment.
200+ tests, CLI, sklearn-compatible
Modular AutoML library for tabular data with preprocessing search, Bayesian optimization, ASHA pruning, ensembling, explainability, and run artifacts.
No-lookahead, cost-adjusted backtests
Walk-forward pairs-trading research framework with cointegration selection, trailing-only features, transaction costs, baselines, and optional PPO evaluation.
Redis rate limiting, health checks, metrics
Config-driven Go API gateway with reverse proxying, load balancing, Redis rate limiting, JWT auth, circuit breakers, retries, and metrics.
Replay dashboard, action logs, fixtures
Autonomous browser agent with perceive-plan-act execution, Playwright automation, session recording, replay dashboard, exports, and human intervention.
Walk-forward eval, no-arbitrage constraints
Hybrid LSTM and GARCH volatility-surface forecaster benchmarked against SABR and Heston baselines with no-arbitrage constraints.
HMAC artifacts, schema sharing
Interactive implied-volatility analytics product backed by a deterministic Python research pipeline and signed JSON artifact bundles.
500+ meters, IsolationForest, risk scoring
Smart-meter anomaly detection prototype for ranking non-technical loss risk with behavioral features and plain-language explanations.
18 tests, parallel mode, highlighting
Grep-like Rust CLI for recursive text search with case-insensitive matching, line numbers, highlighting, and parallel search.
A production-facing ML internship focused on revenue forecasting, sales commission prediction, and scalable batch ML workflows.
Machine Learning Intern
Built and deployed a one-click forecasting pipeline for $150M+ quarterly revenue planning over 1.5M+ records, with interpolation, anomaly detection, smoothing, feature engineering, validation, and batch inference.
Improved commission forecast accuracy by 15% with XGBoost ensembles for 4,900+ sales representatives, reducing MAPE to 6-10% using 60+ lag, trend, seasonality, rolling-stat, and business-calendar features.
Added cold-start logic, quantile forecasts, and market upside/downside scenarios while reducing the end-to-end preprocessing, training, validation, and inference workflow from 6-8 hours to 2 hours.
The stack is organized around the systems Shrey builds: model pipelines, agentic applications, backend services, and deployment infrastructure.