Himanshu Kumar Singh

Senior Data Scientist · AI Engineer

I build LLM systems that hold up in production.

Senior AI engineer with 5+ years across applied ML, backend and data platforms. I work on Text-to-SQL agents, retrieval (RAG), evaluation and the services around them. That work has gone into a government chatbot used by 10M+ people and a 0-to-1 platform with 1,000+ clients. I care more about correctness, observability and evaluation than about demos.

Gurugram, IndiaTredence · ex-Beehyv · ex-LTIMindtreeIIT Ropar

Himanshu Kumar Singh
users served by Kumbh Sahayak
10M+
Text-to-SQL accuracy at GIDR.ai, up from 50%
95%+
clients and leads on a 0-to-1 platform
1,000+
technical interviews conducted
100+

About

Background

I studied Mechanical Engineering at IIT Ropar, where my capstone used LSTMs to replace expensive armour-impact simulations. That got me into ML. Since then I have worked as a data scientist for the Financial Intelligence Unit of India, as a data engineer on US real-estate pipelines, and as a backend and GenAI engineer on products from 0 to 1.

Today I am a Senior Data Scientist at Tredence, building a Text-to-SQL agent for marketing teams at Target. I own the memory, retrieval, evaluation, tracing and latency of that agent. Outside work I build and run small full-stack products on my own infrastructure.

Experience

Where I’ve worked

5+ years across applied AI, backend and data engineering.

  1. Senior Data Scientist · Tredence

    Jul 2026 – Present · Gurugram

    • Building a Text-to-SQL agent (Audience Platform) for marketing teams at Target Corporation.
    • Own its memory management, RAG architecture, evaluation, tracing and latency optimization.
  2. Senior Software Developer · Beehyv Software Solutions

    Aug 2022 – Jul 2026 · Gurugram

    GIDR.ai — Research & production AI

    • Built evaluation and benchmarking for chatbot and Text-to-SQL systems. Raised SQL accuracy from 50% to 95%+ with dynamic few-shot examples mined from query logs.
    • Built a voice-driven B2B workflow platform (React + FastAPI) that turns PDFs and images into structured JSON and drives forms, troubleshooting and guided flows.
    • Built an agent-driven UI (DOM control + voice) on deterministic state-machine workflows, with layout-aware PDF parsing.

    EB1A Experts — Backend, GenAI, DevOps (lead)

    • Led the platform from 0 to 1 in a year: FastAPI services, React front end, CI/CD and observability on AWS. It brought in 1,000+ clients and leads.
    • Built a multi-agent recommendation engine (RAG, web search, MCP). It scores approval odds per USCIS criterion against past successful cases.
    • Cut writers’ manual effort by 70% with generation and NLP-based verification.

    Kumbh Sahayak — Data Scientist

    • Built the multilingual assistant for Kumbh Mela 2025, which served 10M+ users with 10k+ concurrent requests.
    • Reached 90%+ accuracy with llama3:8b using DSPy and TextGrad prompt tuning, fine-tuned BGE embeddings and hybrid retrieval.
    • Added safety guardrails (AWS Guardrails + custom rules), plus a voice-cloning and lip-sync pipeline for official video messages.

    EasyKnock — Senior Data Engineer

    • Built Airflow and DBT pipelines over 40+ sources for business, sales, accounting and marketing teams. Cut the failure rate by 90%.
    • Migrated the ML and Airflow codebase from GCP to AWS, with CI/CD on GitHub Actions.

    BluBracket (acquired by HashiCorp) — Backend Developer

    • Built a software bill of materials (SBOM) module for security analysis and contributed to the multi-task scheduler.

    Internal products — AI Engineer

    • Built deepeval-based evaluation for a multi-tenant agent workflow, and Beecoach, an LLM-driven adaptive assessment tool.
  3. Data Scientist · Financial Intelligence Unit of India · LTIMindtree

    May 2021 – Aug 2022 · Delhi

    • Built a data quality framework for batch reports from 17 types of financial institutions, checked against 32 rules. Includes external API validation against 8 institutions.
    • Worked on ML risk scoring for entities, transactions and locations, peer-group clustering, data models and Neo4j network views.

Selected work

Case studies

Production systems I designed or led, with the numbers that mattered.

Research & coursework

  • Armour design with deep learning

    B.Tech capstone, team of 4. Automated 500 LS-DYNA simulations and trained an LSTM on intermediate time-series states. It beat a feed-forward network in the low-data regime (R² 0.999 vs 0.986, MAPE 2.5% vs 28.6%).

  • Chaotic double-pendulum prediction

    RNN and LSTM models trained on numerically solved trajectories of a chaotic system.

  • Deep learning from scratch

    CNN image classification with Grad-CAM, Hindi→English translation with attention, and RNN, LSTM, WGAN and vanilla GAN implementations.

Toolkit

Skills & credentials

AI / ML
LLM agentsText-to-SQLRAG & hybrid retrievalLLM evaluationDSPy / TextGradEmbedding fine-tuningLangGraphPyTorchTransformersSpeech & lip-sync models
Backend
PythonFastAPIGoJava / Spring BootNode.jsPostgreSQLMongoDBNeo4jKafka
Data
AirflowDBTBigQueryRedshiftETL pipelinesPandasscikit-learn
Cloud & DevOps
AWSGCPDockerKubernetes (EKS)GitHub ActionsKongGrafana / PrometheusJaeger / Loki
Frontend
TypeScriptReactNext.jsExpo

Education

IIT Ropar

B.Tech (Hons), Mechanical Engineering

2017 – 2021 · CGPA 7.61 · Minor in Physics 8.0

Certifications

Contact

Let’s talk

Open to senior and staff roles in applied AI and LLM systems, and to interesting problems in general.