Currently building EviValue at NumericaIQ

Applied AI Engineer, Data Scientist, ML Engineer, AI Enthusiast, Problem Solver

Taljinder
Singh.

M.S. Data Science, ASU '26 · B.Tech, IIT RoorkeeSee projects →

I build LLM-powered products end to end: agentic pipelines, retrieval systems and forecasting models, with the math kept deterministic and every number traceable.

evivalue / budget-impact
01 / 07

What I built at NumericaIQ · hover a step

Research Setup. A plain-English question is rewritten into a structured study question.

Language models do the reading and writing. Deterministic Python does all of the math. The analyst approves each step.

01About

The 60-second version.

Portrait of Taljinder Singh
Taljinder SinghTempe, Arizona

Applied AI engineer and data scientist who builds LLM-powered products end to end.

I completed my M.S. in Data Science at Arizona State University in May 2026, after a B.Tech from IIT Roorkee, and now work as a Junior Applied AI Engineer at NumericaIQ, where I build AI pipelines for health economics on the EviValue platform. Before that, I worked in analytics and pricing strategy at Indus Insights and OYO.

My work spans LLM pipelines, machine learning, forecasting, and optimization. At NumericaIQ, I built a seven-screen AI pipeline that turns a plain-English health economics question into a reviewed cost model and report, with deterministic math and every number traceable to its source. Earlier, my models improved forecasting accuracy, optimized pricing, and drove measurable revenue growth.

Years experience
0+
Years experience
In analytics & ML engineering
Revenue uplift
0%
Revenue uplift
Through pricing optimization
AI pipeline screens
0
AI pipeline screens
Question to report, human-reviewed
Case studies
0
Case studies
Four with live demos

What I'm good at

?questionplanretrievereasontoolsyou approvereport
  • 7-screen LangGraph pipeline with eight Claude-powered steps at NumericaIQ
  • Human approval steps with deterministic math alongside LLMs
  • LLM-as-a-judge relevance ranking and guideline checks
  • LangSmith tracing and automated tests
  • RAG pipelines and QLoRA fine-tuning projects
02Experience

Where I've done the work.

From analytics and pricing to AI pipelines for health economics. Pick a role to read the details.

Current roleFull-timeJun 2026 - Present · Tempe, AZ

Junior Applied AI Engineer

NumericaIQ logoNumericaIQ·EviValue Platform · AI for Health Economics
7
Pipeline Screens
78
Guideline Checks
8
AI-Powered Steps

The role

NumericaIQ builds EviValue, an AI-assisted platform for health economics and outcomes research (HEOR). One of its core analyses is budget impact: an estimate of how much adopting a new drug, test, or device would change a health insurer's or hospital's spending over the next few years. I built EviValue's Budget Impact pipeline end to end, as the platform's sole engineer for its first four months. A user types a research question in plain English, and the pipeline turns it into a study protocol, searches the medical literature, builds a decision-tree cost model, projects costs and savings across a three-year adoption plan, runs sensitivity analysis, and writes the final report. Language models do the reading and writing, deterministic Python does all of the math, and the analyst reviews and approves the work at each step.

Key contributions

    • Designed the end-to-end flow across seven screens: Research Setup, Study Design & Protocol, Model Builder, Model Setup, Results, Sensitivity Analysis, and Report
    • Orchestrated eight Claude-powered steps with LangGraph: question rewriting, protocol drafting, search-query writing, paper relevance screening, evidence extraction, guideline review, report writing, and an in-app assistant
    • Kept every calculation out of the language model: the cost model, budget calculator, validation checks, and sensitivity engine run in deterministic Python, so results are reproducible and traceable
    • Added approval steps so the analyst reviews and approves the protocol, then reviews and can edit every model input before results are produced

Stack

FastAPILangGraphClaude APILangSmithNext.jsPostgreSQLSQLAlchemyAlembicRailwayVercelWebSocketsSentrypytestClerk Auth

Up next · Selected work

Eight case studies on the shelf.

01 · Retail_Forecasting_MLOps_AWS.ipynb

Retail Sales Forecasting & MLOps Platform on AWS

Hover a spine to preview it, click to jump in.

Live demo you can try Click a spine to jump

03Selected work

Things I've built and shipped.

Each project opens into a full case study: the problem, the approach, the architecture, and the results. Flip the preview between the live app, its pipeline, and the code.

01ML & Forecasting Deployed on AWS · now offline

Production-style weekly demand forecasting platform with AWS infrastructure, MLflow governance, and API serving (deployed on AWS, now offline).

Retail MLOps / architecture
Records
421K+
Beats Baseline By
~24%
Deployment
AWS (now offline)
AWSSageMakerMLflowDVCFastAPIDocker
Code
02AI & LLMs Live

QLoRA fine-tuning of LLaMA 3.2-3B on 4,013 Indian recipes with systematic multi-model baseline evaluation, prompt ablation study, MLflow experiment tracking, and a live Gradio demo on HuggingFace Spaces with ZeroGPU (H200).

huggingface.co/spaces/taljindergill78/indian-kitchen
Ingredient F1 Gain
2.85x
Trainable Params
0.75%
Training Time
28 min
PythonLLaMA 3.2-3BQLoRAPEFTHuggingFaceMLflowGradioZeroGPU
Live demo Code
03AI & LLMs Live

Multi-agent retrieval system over payer formulary PDFs with BGE embeddings, cross-encoder reranking, structured LLM extraction, and grounded synthesis. Anton Rx Hackathon, ASU.

rx-radar.vercel.app
Policies Ingested
25+
Vectors
~600
Embedding Dim
768
FastAPIPineconeRAGsentence-transformersGroqNext.jsgrounding
Live demo Code
04AI & LLMs Live

A custom multi-agent system for corporate AML risk assessment. Type a natural language query and get a fully cited risk dossier from 5 live public data sources, with rule-based relevance classification, probabilistic confidence scoring, and a Groq-powered LLM narrative.

web-production-7511d.up.railway.app
Evidence Findings
1,125
Citation Rate
98%
Pipeline Runtime
<3s
PythonFlaskGroqLLaMA 3.1 8BSEC EDGAROFAC SDNCourtListenerGDELTNetworkX
Live demo Code
05AI & LLMs Live

A smart memory coach that doesn't just store what you learn — it tells you what to review, when, and how. Built in 24 hours at Village Hacks Hackathon, ASU.

rewind-mu.vercel.app
API Routes Built
10+
LLM Architecture
Multi-Model
Hackathon Build
24 Hours
Next.jsGroqRAGSupermemoryVercelSendGrid
Live demo Code

Also built

Earlier analytics & ML work

06ML & Forecasting

Built a user-based collaborative filtering recommender on the Book-Crossing dataset (270K+ books, 270K+ users, 1M+ ratings) that suggests five unread books to every reader.

Ratings
1M+
Neighbours per user
10
07Data & Analytics

Built an interactive Tableau dashboard with a team of four that answers 14 HR questions on attrition, pay, performance, and satisfaction.

HR questions
14
Team
4 people
08Data & Analytics

Built a solo PySpark and Spark SQL analysis of the Yelp Open Dataset filtered to Arizona, joining five JSON files (business, review, user, tip, check-in) converted to Parquet on a self-configured Hadoop and Spark virtual machine.

Data files
5 JSON
Engine
Spark SQL

Results · lab notebook

Don't take my word.
Re-run the experiments.

Six questions I had to answer in real projects. Pick one, see the baseline it was measured against, then run it.

Exp 01 · Indian Kitchen

Can a small open model learn Indian cooking?

2.85x

Ingredient F1

Step 1 · Score three untrained models on 500 held-out recipes, then pick one

LLaMA 3.2-1B

0.0991

baseline Ingredient F1

Close, but behind on F1, ROUGE-L and BERTScore

LLaMA 3.2-3B picked

0.1072

baseline Ingredient F1

Best on all three metrics, small enough for QLoRA

Phi-3-mini

0.0615

baseline Ingredient F1

Dropped: degenerate output on 58% of rows

Step 2 · QLoRA fine-tune the winner on 4,013 recipes

Press run to reproduce this result →

Before · base LLaMA 3.2-3B0.1072
After · QLoRA fine-tuned0.3052
See the full story
04Toolkit

What I build with.

80+
Technologies
20+
GenAI & LLM tools
14+
Data engineering tools
10+
Cloud & MLOps tools

The map: tools ↔ where I used them.

Pick a category to light it up, or tap any dot to see the projects and roles behind it.

Hover or tap a dot · drag to play · big circles are projects & roles

05Contact

Let's build something together.

Open to opportunities in Data Science, ML Engineering, and Analytics roles. Whether it's a project, job opportunity, or just a friendly hello, I'd love to hear from you.

Fastest way to reach me

taljindergill78@gmail.com

Send an email