Senior Data Scientist · Credit Risk · Multi-Agent RAG · Recommender Systems
Three ML systems deployed to GCP Cloud Run — built to document failure modes, enforce constraints, and serve live.
4+ years building production ML across credit risk governance, RAG/NLP, and recommender systems.
3 live Cloud Run endpoints · real datasets · every metric reproducible.
Bengaluru, IN
| PulseAgent | PulseGuard | PulseDiscover |
|---|---|---|
| Multi-agent RAG with NLI citation gate | Credit risk governance platform | Hybrid recommender + search IR |
LangGraph Send() fan-out · BM25 + dense RRF · cross-encoder NLI gate enforces ANSWER_WITH_CITATION or ABSTAIN — no hallucination path |
LightGBM_monotonic + 15 directional constraints · Optuna HPO · isotonic calibration · 9-gate governance pipeline · SHAP reason codes | ALS matrix factorization · FAISS · LambdaMART LTR fusion · BM25 + dense hybrid IR · cold-start coverage |
| 56% ANSWER · 44% ABSTAIN · 75.9% NLI precision · 0% error · 218ms latency · 200-query eval | AUC 0.7769 · ECE 0.0034 · KS 0.4141 · 307k real applicants · early-stopping bug caught + fixed | R@20 0.085 (warm) · R@20 0.241 (cold) · FlatIP 47% faster · 0% empty responses |
| Package | What it does |
|---|---|
| metriclens | pip install metriclens · metric movement decomposition — mix shift, rate shift, cross term |
| featureleakagelens | pip install featureleakagelens · pre-training leakage audit reports for tabular ML datasets |
| goldensetauditor | pip install goldensetauditor · LLM/RAG eval dataset auditor — conflicting labels, duplicates, weak references |
| trialcheck | pip install trialcheck · A/B experiment readout auditor — SRM, peeking risk, guardrail movement |
| docingestqa | pip install docingestqa · RAG ingestion pre-check — 11 deterministic checks before indexing |
| inferencelens | pip install inferencelens · inference cost/quality tradeoff auditor — Pareto frontier, routing rules, PASS/WARN/FAIL verdicts |
💼 Open to: Senior DS · Applied AI · Decision Science · ML Governance · Bengaluru (or remote)