"Containerization isn't just deployment—it's a contract between dev and ops for predictable behavior at scale."
"RAG systems fail when retrieval accuracy < 85%. Always benchmark your embeddings before production."
"Explainability isn't optional—SHAP values turn black boxes into trusted production systems."
Driving core platform work at Sylox on a multi-tenant DPSM SaaS — 44 routers, 400+ async REST endpoints, sub-500ms p95 across AKS EKS GKE
Production-grade expertise across the AI systems lifecycle
JD pillars mapped to production evidence — what every line of the role looks like in real shipped systems
Independently designed, shipped, and maintain 44 routers and 400+ async REST endpoints at Sylox; 40+ in the AI Privacy Verifier; 20+ in the matter-isolated RAG. FastAPI + Pydantic v2 + async SQLAlchemy 2.0 across every project, with Flask comfort as a fallback.
asyncpg + uvloop sustaining sub-500ms p95 at 80-concurrency autoscale on a multi-tenant 11-schema Postgres. Celery + Redis + RQ workers with circuit breakers, per-tenant rate budgets, and partitioned tables for bursty load.
LangGraph 7-stage classification pipeline at Sylox with a Claude verifier; LangChain + LlamaIndex RAG over policy documents; provider-agnostic LLM abstraction across 7 vendors (OpenAI, Anthropic, Gemini, Mistral, Cohere, AWS Bedrock, and one cloud-native provider) so a vendor switch is a config flag.
Multi-stage Dockerfiles (distroless runtime, non-root UID, <180MB images); Helm chart with values-{dev,stage,prod}.yaml, HPAs, PodDisruptionBudgets, NetworkPolicies, pre-rendered ServiceMonitors. Docker Compose for local parity with make demo bootstrapping the full stack in <90s.
Strict TDD with red-tests-first commits — pytest + pytest-asyncio, testcontainers (real Postgres + Redis, no mocks), Hypothesis property tests at 500+ examples, Playwright + Vitest E2E, schemathesis contract tests, k6 + Locust performance harnesses; gated by diff-cover ≥ 80% on every PR.
Owned multi-stage pipelines (build → test → security-scan → publish → cluster deploy via Helm) with environment approvals, blue/green rollout, and automatic rollback on failed health gates. Cut PR-to-deploy from ~45min → ~9min via dependency caching, test sharding, and BuildKit cache mounts. Same patterns shipped on Azure DevOps and GCP Cloud Build.
OAuth2 + OIDC with Auth0 (pluggable to Okta or Azure AD via the same provider interface), JWT (python-jose), Argon2id hashing, refresh-token rotation, per-endpoint RBAC via require_permission(). AKS workload identity, NetworkPolicies default-deny, Pod Security Standards: restricted, Key Vault CSI mount. SCIM 2.0 provisioning at Sylox.
Bash + PowerShell automation for stack bootstrap, evidence-pack export, and on-call diagnostics; make demo and make preflight orchestrate end-to-end environment up + canary→PDF pipeline assertion in under 90 seconds.
Production GenAI work across SaaS platforms, contracts, and independent research
Driving core platform engineering on IRIS — a multi-tenant Data Privacy & Security Management SaaS for DPDP Act 2025, GDPR, HIPAA, RBI, and SEBI compliance. 44 routers, 400+ async REST endpoints, 11-schema Postgres, sub-500ms p95 at 80-concurrency. Architected a 7-stage LangGraph classification pipeline (regex → Presidio + spaCy → Claude verifier → analyst override) lifting sensitive-data detection accuracy 18% across 10K+ records. Productionized cloud-portable Terraform deploys across GCP Cloud Run, AKS, and EKS. Operationalized GCP Cloud Build CI/CD with a 30-gate verification system and 161 idempotent Alembic migrations.
Architected a multi-tenant Pinecone × Claude RAG over 5K+ legal-matter documents, eliminating cross-tenant context leakage via strict namespace isolation. Async ingestion pipeline processed 1GB+ batches across direct upload, n8n webhooks, S3 async, and ZIP batch — cut redundant embeddings 35% and embedding cost 30% via SHA-256 dedup and 50-per-request batching. Lifted retrieval relevance 28% with hybrid semantic + BM25 + cross-encoder reranking; query latency < 2s p95. Shipped 20+ FastAPI endpoints on AWS (EC2, S3, RDS, Redis) behind NGINX.
Used the post-graduation runway to deepen production GenAI specialization. Earned nine industry certifications in eight weeks across GenAI, Vector Search, MLOps, Kubernetes, and Apache Airflow — Oracle OCI 2025 (5 certs), Databricks (2), Astronomer Apache Airflow 3 (2), KodeKloud (RAG + K8s). Took freelance fine-tuning and RAG-prototyping engagements; built open-source side projects including Cerevra (ContextOS) and early prototypes of Sentinel.
Production-grade AI systems deployed at scale
9-service event-driven backend that continuously probes AI provider APIs and produces RFC 3161 trusted-timestamped evidence packets — canary traps, membership inference probes, model fingerprinting, policy surveillance via Azure Document Intelligence. 40+ FastAPI endpoints, provider-agnostic LLM layer over 7 vendors, Helm-deployed on AKS with Azure DevOps blue/green pipelines.
Multi-tenant retrieval over 5K+ legal-matter documents. Pushed isolation into the vector DB via per-matter Pinecone namespaces — enforced before retrieval, not after. Async ingestion across upload, n8n, S3, and ZIP batch (1GB+); SHA-256 dedup cut embedding cost 30%; hybrid BM25 + reranking lifted relevance 28% at < 2s p95.
Production-grade RAG engine that treats retrieval like a database — write-ahead logging with fsync, deterministic crash recovery, hardware-grade tenant isolation, no vector store dependency. BM25-first (no GPU), 4-dim Context Quality Scoring, citations anchored to source offsets. Docker, K8s manifests, Python + TypeScript SDKs. 219 commits.
LangGraph-orchestrated multi-agent system: a router dispatches to specialized Claude and Codex agents (diff analysis, secret scan, dependency audit, playbook retrieval), each with tool-level RBAC. Async FastAPI webhooks consume GitHub PR events; long reviews run on Kubernetes worker pods deployed via a custom Helm chart, with full audit trail of every tool invocation.
Final-year capstone (Team Lead) — fine-tuned Vision Transformer for real-time suspicious-activity detection from user-controlled camera APIs. Three-stage pipeline with WebSocket / e-mail / SMS alerting. Trained on UCF-Crime + KTH Action; 95% manual-audit alert precision. Published in TIJER (May 2025).
Peer-reviewed research and long-form technical writing
Vision Transformer–based real-time suspicious-activity recognition from user-controlled camera APIs. 92% detection accuracy on UCF-Crime, 100ms / frame, 95% manual-audit alert precision. Authors: Ashwin Upadhyay, Sarrah Burmawala, Amit Bhosale (TCOER, SPPU).
Read the PaperA career-narrative piece on building cross-disciplinary depth — from HTML/CSS/JS and full-stack web foundations into SQL, R, Python, ML, and neural networks.
Read on MediumA primer on the five foundational ML models — Linear Regression, Logistic Regression, Decision Trees, Random Forests, and Gradient Boosting — framed for business audiences.
Read on MediumIndustry-recognized credentials validating production-scale expertise
Databricks
Databricks
Astronomer
Astronomer
Oracle Cloud Infrastructure
Oracle Cloud Infrastructure
Oracle Cloud Infrastructure
Oracle Cloud Infrastructure
Oracle Cloud Infrastructure
Calyptus
Securiti
Saviynt
KodeKloud
KodeKloud
Oracle Cloud Infrastructure
HackerRank
Professional Certifications
Active Credentials
2025–26 Certifications
Production impact, recognition, and competitive milestones
500+ problems solved with a 500-day submission streak. 100 Days Badge ’24 / ’25 / ’26 · 200 Days Badge ’25 · Annual Badge 2025 · monthly badges April–November 2025 plus February 2026.
Earned the AI Fluent Tech Professional credential in the top 5% of the global Calyptus cohort, validated across Agentic AI, Explainable AI, and GenAI competencies.
Open-source maintainer of Cerevra — 219 commits, 2,195 tests, 80% diff-coverage. GitHub badges: Pair Extraordinaire · Pull Shark ×2 · YOLO · Quickdraw across 36 public repositories.
Lifted production retrieval accuracy to 90% across AI-driven systems; reduced end-to-end ML pipeline latency 40% through async I/O, caching, and batching. Eliminated cross-tenant leakage in two multi-tenant platforms via vector-DB-level namespace isolation and OIDC-bound RBAC.
Team Lead on the SHAR (Scalable Human Activity Recognition) capstone — peer-reviewed publication in TIJER (ISSN 2349-9249, Volume 12, Issue 5).
Continuous contributions and open-source engagement
Commits This Year
Pull Requests
Total Stars
Repositories
Real-time insights and visitor metrics
Load Time
0.8s
First Paint
0.3s
Lighthouse Score
98/100
Open to AI/ML Engineer · GenAI Engineer · Backend AI Platform roles — Hybrid / Remote / Onsite (Pune, Bengaluru, Hyderabad)
ashwinupadhyay09@gmail.com
+91 8329396282
Pune, Maharashtra, India 411046
B.E. Information Technology · TCOER, SPPU · CGPA 8.53 / 10 · 2021–2025
Request sent successfully!