IAP · GenAI / SAP · 2025–
Commercial Intelligence OS
Billing NLP in production inside SAP Joule workflows.
Bedrock · SAP · Python
system boot · null∅verflow · crt phosphor
Build, ship, and harden intelligent systems.
AI/ML studio led by Rudhvish Patel — MS AI (DePaul), AI/ML Engineer Technical Lead. Production GenAI, ML platforms, and research that survives contact with reality. Inquiry only. No public pricing.
Work
RUSH / DePaul · Research · 2025
Syndemics · IEEE eScience
Ensemble clustering published at IEEE eScience 2025.
PyTorch · scikit · HPC
ReDDDoT · LLM · 2024–
Multilingual LLM + recs
Multilingual gen + ranking tuned for live journeys.
Transformers · FastAPI
Techify · ML platform · 2023–
SageMaker lead ML
Lead ML + voice follow-up closed the score loop.
SageMaker · Lambda
CV / OCR · Systems
Road-sign vision pipeline
Detection + OCR shipped as one AWS system.
YOLO · OCR · AWS
Education · Recsys
Education video recommender
Ranking that respects pedagogy, not just clicks.
Recsys · Feature store
Capabilities
> cat /capabilities/*.term — problem · approach · deliverable · stack
null∅verflow:/capabilities/01.termPHOSPHOR · OK
01
LLM & RAG systems
- > PROBLEM
- Enterprise knowledge is trapped in silos; chat demos collapse under permissions, latency, and hallucination risk.
- > APPROACH
- Ground Claude / Bedrock / Mistral with retrieval over governed corpora, tool-use contracts, and eval gates before any user-facing path ships.
- > DELIVERABLE
- Production RAG + agent surfaces with citation trails, rate limits, and rollback — not slideware demos.
stack_tags[]
null∅verflow:/capabilities/02.termPHOSPHOR · OK
02
ML engineering
- > PROBLEM
- Models that look good in notebooks fail under drift, cost spikes, and missing ops once real traffic arrives.
- > APPROACH
- Train, evaluate, and deploy on SageMaker/AWS with feature contracts, shadow traffic, and monitoring that pages humans — not dashboards nobody reads.
- > DELIVERABLE
- Endpoints with SLOs, cost envelopes, and retrain loops that survive production load.
stack_tags[]
null∅verflow:/capabilities/03.termPHOSPHOR · OK
03
Data & analytics platforms
- > PROBLEM
- Models starve when pipelines are brittle, schemas drift, and SAP/analytics truth is untrusted.
- > APPROACH
- Build validated ingestion, quality gates, and analytics stacks that keep downstream ML fed with reproducible, auditable data.
- > DELIVERABLE
- Pipelines + marts with lineage, freshness SLAs, and clear ownership — so models stop guessing.
stack_tags[]
null∅verflow:/capabilities/04.termPHOSPHOR · OK
04
Computer vision / OCR
- > PROBLEM
- Detection and document/sign pipelines stall between dataset design and shippable inference.
- > APPROACH
- Own the loop: labeling strategy, model selection (YOLO/OCR), latency budgets, and AWS packaging as one system.
- > DELIVERABLE
- Inference services with measured accuracy, batch + online paths, and ops hooks for bad frames.
stack_tags[]
null∅verflow:/capabilities/05.termPHOSPHOR · OK
05
Applied research
- > PROBLEM
- Papers and prototypes die in the lab when product constraints never enter the experiment design.
- > APPROACH
- Design studies with transfer criteria up front — metrics, compute envelopes, and failure modes that map to a shippable surface.
- > DELIVERABLE
- Published or internal research plus a prototype→product handoff path when the science has to land.
stack_tags[]
null∅verflow:/capabilities/06.termPHOSPHOR · OK
06
Automation & agents
- > PROBLEM
- Manual ops burn cycles; naive agents amplify risk when AppSec is bolted on after the demo.
- > APPROACH
- Replace brittle workflows with constrained agents — least privilege, vault secrets, human checkpoints, and audit logs from day one.
- > DELIVERABLE
- Agentic ops that cut toil without opening the blast radius — measurable time saved, not vibes.
stack_tags[]
Stack wall
Python
PyTorch
TensorFlow
scikit-learn
Hugging Face
AWS SageMaker
Bedrock
Lambda
S3
FastAPI
Next.js
TypeScript
Prisma
Postgres
Docker
RAG
LangChain
OpenCV
Spark
Airflow
Lab
lab/01 · live
Scale Arena
Two objects, one question: which is physically larger? Daily deck, amber score.
lab/02 · live
MAGNITUDE RUN
Drag the amber silhouette to match the phosphor reference across five daily rounds.
lab/03 · live
Eval harness
Grade 10 RAG and classifier outputs against a rubric. Precision, recall, agreement /100.
lab/04 · live
Prompt forge
Snap role, rule and example blocks together, run 8 scripted tests, diff each version.
lab/05 · live
Secure ops
Spend a small budget on rate limits, auth gates, CSP and a secrets vault. Block abuse, not users.
lab/06 · live
Gradient Descent Golf
Tune learning rate and momentum, shoot, and watch real gradient steps roll toward the hole.
lab/07 · live
Overfit or Not
Drag polynomial degree 1–12 and watch train and validation error pull apart. Lock the sweet spot.
lab/08 · live
Attention Lens
Hover words to light attention arcs, then call which word 'it' attends to most.
About
Fremont-based AI engineer. Focused on production GenAI/ML that businesses can actually run — not slideware. MS in Computer Science (AI), DePaul; AI/ML Engineer — Technical Lead.
AWS AI PractitionerAWS CloudSecurity+IEEE eScience 2025