Plain-text version
- Who: ML & AI Engineer · MS Computer Science, NJIT · NYC metro, open to relocating anywhere in the US
- Focus: production ML · LLM agents & RAG · causal inference · deep learning
- What I bring to a team: I take a problem from raw data to a model, an API and a live deployment, and I audit my own work. I found 12 defects in LoanLens, a label leak in SessionScout and a biased "ground truth" in a 14M-row benchmark. Then I fixed them and added tests.
- Looking for: ML Engineer · AI Engineer · GenAI Engineer · Forward Deployed Engineer roles, startups included
- Contact: mohammed.ahmeduddin16@gmail.com
| 🎯 Best test AUC | ⚡ Latency | 📊 Largest dataset | 🏥 Real records | ✅ Tests | 📦 Live deployments |
|---|---|---|---|---|---|
| 0.958 | 27 ms | 14M rows | 39,694 | 623 | 4 |
| SessionScout LSTM | Redis-cached API | LiftLab · Criteo | CareAgent · CMS providers | across 4 ML projects | Hugging Face + GCP |
A LangGraph supervisor routes 5 agents over 39,694 real New Jersey Medicare providers (CMS 2024, $4.78B in payments).
- Peer comparison: providers are compared within their specialty (71 specialties) after adjusting for patient risk. It flags 1,461 outliers (3.7%), each with reasons drawn from the data, and an independent Isolation Forest agrees (outliers sit at the 96th percentile of its score vs the 48th).
- Checked LLM summaries: every number in an LLM summary is checked against the data in code.
- Caching: a fingerprinted cache cuts repeat requests from 4.7 s to 0.83 s with 0 tokens.
LangGraph OpenAI Isolation Forest FastAPI PostgreSQL GCP Cloud Run Secret Manager
An XGBoost credit-risk model on 307,511 loan applications (16M+ joined rows).
- Calibrated: predicted default rate 0.080 vs actual 0.081.
- Fair-lending checks: sex, marital status and age were removed at a cost of 0.004 AUC, and the four-fifths rule check passes (0.95 vs the 0.80 threshold).
- Explanations: SHAP values become CFPB adverse-action reason codes, retrieved by RAG over 9,977 CFPB regulation chunks. Every citation and reason is verified in code.
- Self-audit: found and fixed 12 defects.
XGBoost SHAP RAG ChromaDB OpenAI dbt PostgreSQL FastAPI MLflow
A 4-model cascade (Logistic Regression → XGBoost → LSTM → Transformer) in PyTorch on 241K sessions. The LSTM reaches 0.958 test AUC.
- Label leak found: live-demo testing exposed a leak that marked 5,640 buyer sessions. The pipeline was rebuilt and retrained, and regression tests now guard against it.
- Serving: FastAPI with Redis caching in Docker responds in 27 ms.
PyTorch LSTM Transformer XGBoost FastAPI Redis Docker HuggingFace
Measures how much ads caused visits and conversions on the full Criteo Uplift dataset (13,979,592 users).
- Randomization check: the dataset isn't fully randomized. A classifier predicts who was treated (AUC 0.510, vs at most 0.5014 with shuffled labels), so the raw comparison overstates the visit effect by 39%. Cross-fitted AIPW is used as the reference instead.
- Targeting: one quarter of users produce ~94% of the extra conversions. A Causal Forest ranks users with a held-out Qini of 0.339.
- Validated first: every estimator is checked on simulated data with a known answer before it touches real data.
EconML DoWhy AIPW Causal Forest scikit-learn FastAPI Streamlit
| Practice | Example from my work |
|---|---|
| Audit before claiming | LoanLens: an inverted SHAP sign made every denial list the applicant's strengths; found, fixed, regression-tested |
| Calibrate, then decide | Replacing inflated scores with calibrated ones cut applicants sent away from auto-approval from 46% to 8.7% |
| Verify LLM output in code | LoanLens checks every regulation the LLM cites against the retrieved passages; the decision still works if the LLM is down |
| Question the data | LiftLab: a randomization check on 14M rows showed the "ground truth" itself was biased |
| Honest metrics | CareAgent: renamed a "quality score" to a billing-pattern comparison, because Medicare billing data has no outcomes |
| Bound LLM cost and latency | CareAgent reuses summaries keyed by a SHA-256 fingerprint of model + prompt + facts, with a 45 s time budget and per-client rate limits |
| Ship with CI | Lint, tests and coverage gates on every pull request |


