GenAI / RAG
Kubernetes Gated RAG
Kubernetes Q&A that answers only from the docs you ingest. A gate guards every stage, and infrastructure failures return an uncached outage message, not a guess.
ML and LLM systems built to hold up under scrutiny.
AI/ML engineer building recommenders, RAG, agents, MLOps, and causal inference systems. Each project ships with tests, benchmarks, and documented limits.
42/50planted sales spikes and drops caught by the best detectorplanted sales glitches caught4.8/5AI research brief completeness, up from 3.8 with self-checkingresearch score, up from 3.81%needless model rebuilds on unchanged data, down from 40%needless rebuilds, down from 40%5xthe profit of random outreach in a simulated retention campaignmore profit than random, simulated
I write the evaluation before the model.
I build ML and LLM systems end to end: retrieval, agents, recommenders, causal inference, and the MLOps around them. I report what the measurements show, including where the system falls short. Everything here runs on free-tier infrastructure.
- Focus
- LLM systems, evaluation, MLOps
- Education
- B.Tech CSE, Lovely Professional University
- Location
- Varanasi, India
- Availability
- Open to remote or relocation
Selected projects
Three of nine case studies. Each one lists what its evidence does not show.
AI Agents / MCP
IssueOps MCP
A GitHub issue triage system where the model can only propose. Nothing reaches GitHub until a person approves it, and every step is audit-logged.
MLOps
Continuous Credit Risk Governance Pipeline
A credit-risk model that runs its own lifecycle: it detects drift, retrains a challenger, gates promotion with a bootstrap test, and rolls back if the old model wins.
Skills
Modeling, LLM systems, and the production tooling around them.
Languages
PythonSQLBashC++
Data & Applications
FastAPIStreamlitPostgres (Neon)SupabaseSQLitepandasNumPyTableauPlotly
Statistics & Causal Inference
Power AnalysisWelch's t-testBenjamini-HochbergRosenbaum BoundsIPW / AIPWCATE (T-learner)QiniMcNemar / Bootstrap CIstatsmodelsscikit-uplift
GenAI & Agents
RAGLangGraphMulti-Agent OrchestrationMCP (stdio server)QdrantFAISSFlashRankRagasNeMoGuardGroqNVIDIA NIMGemini API
MLOps & Production
MLflowDVCEvidentlyDrift Detection (K-S)Model Governance & RollbackGitHub ActionsCI/CDLogfireDockerAWS
Machine Learning
XGBoostLightGBMscikit-learnPyTorchProphetSARIMAIsolation ForestALS / BPRTwo-Tower & SASRecSHAPOptuna
Bold marks the tools used most across the projects.
Experience & education
Where the work happened, and the record behind it.
Experience
Data Science Intern
Coding Blocks
- Built an XGBoost customer-churn model and tuned it with GridSearchCV, landing 9 percent above the baseline accuracy the team had been working from.
- Scraped and cleaned over 5,000 competitor reviews into a structured dataset the product side could query directly.
- Covered the full loop: ingestion, feature work, model selection, evaluation, and handover.
Recognition
- Dean's List, top 10 percent of students, 2023, 2024, and 2025
- 5-Star Badge in Python and C++ on HackerRank
Education
B.Tech, Computer Science and Engineering
Lovely Professional University, Punjab, India
Certifications
- SAP Certified, Data Analyst, SAP Analytics Cloud
- IBM DevOps & Software Engineering Professional Certificate
- NPTEL Cloud Computing
- Google IT Automation with Python
Open to AI/ML engineering & Data Science roles.
If you are building something that needs a system tested as hard as it is trained, I would like to hear about it.
Back to home
All projects
Nine end-to-end projects across LLM systems, MLOps, and applied statistics. Each ships with tests, an evaluation approach, and a written list of what it does not prove.
Showing 9 of 9 projects
GenAI / RAG
Kubernetes Gated RAG
Kubernetes Q&A that answers only from the docs you ingest. A gate guards every stage, and infrastructure failures return an uncached outage message, not a guess.
0.95faithfulness on the 8-row starter set
3generation failover links, same model
2cache layers, exact and semantic
AI Agents
Competitive Intelligence Agent
A research agent that turns a company name into a sourced brief. A Critic loop fills gaps, and anything unsourced is marked insufficient information.
+1.0completeness gain from the Critic loop (0 to 5, n=5)
+0.8groundedness gain from the Critic loop (0 to 5, n=5)
3max replan cycles
AI Agents / MCP
IssueOps MCP
A GitHub issue triage system where the model can only propose. Nothing reaches GitHub until a person approves it, and every step is audit-logged.
10MCP tools: 5 read, 5 propose
0execute tools exposed to the model
48hproposal expiry window
MLOps
Continuous Credit Risk Governance Pipeline
A credit-risk model that runs its own lifecycle: it detects drift, retrains a challenger, gates promotion with a bootstrap test, and rolls back if the old model wins.
1%false retrains on undrifted batches (was about 40%)
3promotion checks: tolerance, dominance, significance
105automated tests
Predictive Modeling
Customer Churn Prediction with Profit Optimization
Predicts 90-day churn and targets a customer only when a retention offer's expected profit exceeds its cost, not by a raw probability cutoff.
1.07Mraw transaction records
12RFM features per window
70offline unit tests
Forecasting
DemandLens: Retail Demand Forecasting and Anomaly Benchmark
Benchmarks four forecasting models on 60 retail series, scores two anomaly detectors on injected anomalies, and verifies every number in an LLM-written report.
60retail series benchmarked
0.974best holdout MASE, LightGBM, within 0.04 of two rivals
52automated tests
Recommender Systems
ReelBench: Multi-Stage Recommender Benchmark
Five recommenders scored head to head on MovieLens 25M through one leakage-tested harness, trained on free-tier GPUs and served on CPU.
0.0485Recall@10 for SASRec + ranker, against 0.0243 for popularity
5approaches compared
0external calls at request time
Causal Inference
Causal Impact & Heterogeneous Response Analysis
Tests four causal estimators against a reference effect on 14 million randomized ad rows, then checks which audiences responded and whether that survives multiple testing and hidden bias.
+417%naive OLS bias at strong confounding
0.0565Qini coefficient, 95% CI 0.030 to 0.085
1.12critical Gamma, fragile to hidden bias
Experimentation
Experiment Design & Power Analysis for a Shipping Policy Change
A preregistered A/B pipeline: metric, minimum lift, sample size, and tests are frozen before any data exists, then checked against an effect of known size.
2,503sellers per arm required
2.0ppcomplaint-rate guardrail margin, frozen up front
71automated tests
Interested in working together?
Contact me