I'm a Data Scientist at Softeon, working on production multi-tenant RAG and conversational AI for enterprise supply chain software, while finishing my BS at IIT Madras. My focus is on the engineering side: making GenAI reliable, observable, and cost-predictable in systems real users depend on.
Before Softeon, I shipped a GPT-powered LinkedIn outreach system at Second Brain Labs and taught Python and ML to 100+ students at Antern. Building things and explaining them clearly have both been part of the work from the start.
Role → Data Scientist · Softeon, Chennai (Full-time)
College → IIT Madras · BS Data Science · Expected Nov–Dec 2026
Focus → Production RAG · Agentic AI · GenAI System Design
Open to → Remote GenAI/ML roles · US/EU timezones
- Converted internship → full-time at Softeon while still in college; end-to-end ownership of production multi-tenant RAG pipelines serving real enterprise customers
- Built a kapa.ai-inspired multi-tenant RAG platform from scratch - 10 Docker containers, hybrid search, MCP server, Prometheus/Grafana observability
- Cross-encoder reranking improved Context Precision by +0.15 (measured with RAGAS); grounding validation runs as a blocking step
- RAGAS eval tied to CI/CD with auto-rollback on quality drop; golden query set per tenant, nightly + per-ingestion runs
- IIT Madras Topper Badges in Python, Bash, ML (Rank 106 / 1700+, Score 93/100)
- Mentored 100+ students in Python & ML · Launched two free cohort-based courses (PY001 & PY002)
The problem: Developer-tool companies use products like kapa.ai to power AI assistants over their docs, GitHub repos, and PDFs. I wanted to understand what it actually takes to build something like this with real multi-tenancy and observability, so I built it.
What I shipped: A production-grade, multi-tenant RAG platform across 10 Docker containers, with an MCP server that exposes the full pipeline as a native tool for Claude Desktop.
| Layer | What it does |
|---|---|
| Ingestion | Docs (BeautifulSoup + HeadingAwareChunker), GitHub repos (AST-based code chunker), PDFs (pymupdf4llm) → Celery async workers |
| Query pipeline | SHA-256 cache → hybrid search (RRF fusion of dense + sparse) → Cohere reranker (top-20 → top-5) → GPT-4o-mini via SSE stream |
| Multi-tenancy | Separate Qdrant collection per tenant · Redis cache keyed by sha256(tenant_id + query) · API keys stored as SHA-256 hash only |
| Freshness | HMAC-verified GitHub webhooks (~10s incremental vs ~8min full re-index) · 6h Celery Beat polling · atomic S3 + DB cleanup on delete |
| MCP Server | search_knowledge_base (full pipeline) + fetch_and_query_online_docs (ephemeral, zero Qdrant writes) — stdio + SSE transport |
| Observability | Prometheus + Grafana · LangSmith traces · RAGAS eval (faithfulness + context precision per source type) |
Architecture:
Key decisions and why:
- RRF over weighted sum — rank-based fusion avoids calibrating incomparable dense/sparse score scales
- Per-tenant Qdrant collections over shared + filter — hard isolation, zero query overhead, independent scaling
acks_late=Trueon Celery tasks — task stays on queue until ACK; no silent data loss if a worker crashes mid-job
What I'd do differently: Proper React frontend instead of Streamlit, and per-tenant cost dashboards built in from day one.
FastAPI Qdrant Redis PostgreSQL Celery OpenAI Cohere FastMCP Docker RAGAS Streamlit
Enterprise RAG powering conversational AI for supply chain software, used by real customers.
- Pinecone namespace isolation per tenant — no shared collection, no filter overhead, independent scaling per client
- Cross-encoder reranking (top-10 → top-3); inline grounding validation as a blocking step before any response is returned
- RAGAS eval on nightly + per-ingestion runs; golden query set tied to CI/CD with auto-rollback on quality drop
- Circuit breakers + fallback LLM routing; context drift and embedding distribution shift detection
- Stack: OpenAI · Anthropic Claude · AWS Bedrock · FastAPI · AWS (EC2, Lambda, DynamoDB, SQS, Cognito, ECR, CloudWatch)
GPT-powered outreach system integrated with the LinkedIn API. Handled live campaign traffic across multiple client accounts — automated lead qualification and multi-turn conversation flows.
GPT-4 LinkedIn API Python
Older Projects
End-to-end MLOps pipeline: data ingestion → training → deployment on AWS EC2.
MongoDB Docker FastAPI AWS EC2 CI/CD
Real-time fraud detection pipeline with full MLOps instrumentation and alerting.
AWS Kubernetes Prometheus Grafana DVC MLflow Dagshub
ML-driven campaign optimization with an A/B testing framework for maximizing click-through rates.
GenAI / RAG
Vector Databases & Storage
Cloud & Infrastructure
Languages & ML
Joined IIT Madras in 2022 with one goal: get hired in industry before graduating, without relying on campus placements. Spent the first six months stuck in a Python tutorial loop — kept re-learning the basics without shipping anything. Breaking out of that by doing real projects changed the trajectory.
First internship was at Second Brain Labs in Sep 2024, shipping a production chatbot. Took on ML teaching at Antern at the same time. Joined Softeon as a data science intern in May 2025, converted to full-time by August — while still two years from graduation.
I write about what I've actually shipped — production RAG failures, multi-tenancy trade-offs, GenAI system design — on LinkedIn and Medium.
- ✍️ LinkedIn — RAG failures, eval pipelines, AI NFRs, career lessons
- 📝 Dev.to — technical deep-dives
- 🎥 YouTube — ML content
- 👨🏫 Mentored 100+ students · PY001 & PY002 free cohort-based Python courses
Open to remote GenAI/ML engineering roles (US/EU timezones). If you're building production AI systems or just want to talk shop about RAG/agents, feel free to reach out.
