Hi, I'm Vishnu Abarajithan

The problem

A knowledge-heavy team was losing hours every week digging through PDFs, spreadsheets, scanned images and SQL exports to answer routine questions. Off-the-shelf chatbots returned confident but unverifiable answers — a non-starter for decisions that had to be auditable.

The solution

• Designed an agentic RAG pipeline with LangGraph that analyses each query, runs hybrid retrieval — dense embeddings fused with a BM25 sparse index via Reciprocal Rank Fusion in Qdrant — then cross-encoder reranks before answering.
• Made every response grounded with citations, and added a self-checking retry loop that re-queries automatically when retrieval signals are weak, so the system fails safe instead of hallucinating.
• Shipped it as a FastAPI service with auto-generated Swagger docs and a build-free chat UI in pure Python (FastHTML), instrumented end-to-end with MLflow tracing for latency, token cost and retrieval quality.

Business impact

• Answer lookup dropped from hours of manual searching to seconds.
• Every answer is citation-backed and auditable — safe for real decisions.
• Full tracing means issues are diagnosed in minutes instead of guessed at.

Tech stack

• LangGraph
• Google Gemini
• Qdrant
• BM25 + RRF
• FastAPI
• FastHTML
• MLflow

The problem

A scaling product was spending thousands a month on a closed-model API, with per-token pricing that grew with every new user — and sensitive data leaving their environment on every request.

The solution

• Benchmarked open-weight models and deployed a quantized Qwen2.5 model with vLLM on a GPU VPS, exposed through an OpenAI-compatible endpoint so the client's existing code worked unchanged.
• Containerised with Docker behind Nginx with TLS, request batching and autoscaling, and added Prometheus + Grafana for throughput, latency and GPU utilisation.
• Tuned quantisation, context length and concurrency to hit the target latency at a fixed, predictable monthly cost.

Business impact

• ~70% reduction in inference cost versus the previous API bill.
• All inference runs in the client's own environment — data never leaves it.
• A flat, predictable cost instead of usage-based bills that scaled with growth.

Tech stack

• vLLM
• Qwen2.5
• Docker
• Nginx
• GPU VPS
• Prometheus
• Grafana

The problem

An operations team was manually triaging inbound leads and support emails, copying data between tools and compiling reports by hand — slow, error-prone, and impossible to scale with headcount alone.

The solution

• Built an AI agent on n8n that reads incoming messages, classifies intent with an LLM, drafts context-aware replies and routes each item to the right person or system.
• Connected CRM, email, Slack and a Postgres store, with a human-in-the-loop approval step for anything sensitive.
• Added scheduled workflows that compile and deliver daily reports automatically.

Business impact

• Around 15 hours of manual work removed per week.
• Response times cut from hours to minutes.
• Fewer hand-off errors, and a process that scales without adding headcount.

Tech stack

• n8n
• OpenAI
• Postgres
• Webhooks
• Slack API
• Gmail API

The problem

A growing business was fielding the same customer questions over Telegram all day. Replies were limited to working hours and staff spent time re-typing the same answers — slow, inconsistent responses that cost the business leads and goodwill.

The solution

• Built a conversational AI agent on n8n that receives Telegram messages through the Bot API, classifies intent with an LLM, and replies instantly with grounded, on-brand answers drawn from the business's own FAQ and product information.
• Added conversation memory so the bot follows multi-message threads, plus a human-in-the-loop handoff that escalates anything sensitive or unresolved to a staff member — with the full conversation context attached.
• Instrumented the flow with logging and lightweight analytics so the team can see what customers ask most, and kept everything as maintainable visual n8n workflows the client can extend without code.

Business impact

• Customers get accurate answers in seconds, around the clock — not just office hours.
• Repetitive questions are handled automatically, freeing staff for high-value conversations.
• Consistent, on-brand replies with a clean escalation path for anything the bot can't close.

Tech stack

• n8n
• Telegram Bot API
• OpenAI
• Vector store
• Webhooks

The problem

Years of business logic were locked in legacy SQL Server stored procedures — expensive to run, hard to maintain, and a major blocker to modern analytics. A manual migration would take months and risk subtle logic errors.

The solution

• Built a GenAI translation engine (OpenAI API + few-shot prompting) that converts complex T-SQL — handling dialect nuances like DATEADD and OUTER APPLY — into optimised Databricks SQL.
• Auto-generated dbt models from the translated logic with Python, and programmatically opened validated pull requests into the main branch.
• Wired the whole thing into CI/CD so migration became a repeatable pipeline, not a one-off project.

Business impact

• Migration effort cut from months of manual rewriting to a repeatable pipeline.
• Consistent, reviewable output — every change ships as a validated PR.
• A modern Lakehouse foundation unlocked for downstream analytics and AI.

Tech stack

• Databricks
• Python
• OpenAI API
• dbt
• T-SQL
• CI/CD

The problem

Rich proprietary survey data wasn't translating into decisions. Leaders lacked a clear, trusted view of what was driving customer loyalty, and data preparation was slow and manual.

The solution

• Built robust Alteryx ETL workflows to cleanse, standardise and prepare large survey datasets reliably.
• Designed high-visibility Tableau dashboards that surface NPS drivers and benchmarks for the NPS Prism product.
• Optimised dashboard performance and managed enterprise permissions for a secure, organisation-wide rollout.

Business impact

• Leaders get a single, trusted view of the drivers behind customer loyalty.
• Slow manual data prep replaced with repeatable, reliable pipelines.
• Faster, data-driven decisions on where to improve customer experience.

Tech stack

• Tableau
• Alteryx
• Survey data
• ETL