About
About Eeshan Gupta
AI Engineer building GenAI, RAG and voice AI systems.
Authoring speed
GenAI document drafting platform for a top pharmaceutical client
Codification effort cut
NLP extraction pipeline
Ticket volume cut
RAG-based customer support system for a leading global travel technology provider
Review speed
NLP relevance & classification pipeline for legal e-discovery
Experience
Since 2019, building AI that has to work the first time
From regulated pharma document generation to real-time speech pipelines.
- Shipped a trigger-based call-summarization plugin on GPT models, automating post-call review for customer-facing teams.
- Built a production speech-to-text plugin for real-time call transcription using Whisper and ElevenLabs ASR.
- Benchmarked STT/ASR providers — OpenAI, ElevenLabs, Google — on accuracy, latency, and cost to guide production vendor decisions.
- Designed an LLM-driven packaging recommendation system for a logistics client, improving cost efficiency at scale.
- Built a RAG-based customer support system for a leading global travel technology provider, cutting support ticket volume by up to 50%.
- Evaluated agent frameworks (Crew.ai, Agno) to orchestrate autonomous data-analysis pipelines.
- Architected a GenAI document-drafting platform for a top pharmaceutical client — 5x faster authoring, no technical skills required to operate.
- Built an AI-assisted patent drafting tool that cut authoring time 3x and sped up IP filing cycles.
- Delivered a large-scale NLP extraction pipeline that reduced manual codification effort by 90%.
- Shipped semantic search and RAG across legal, sales, and marketing repositories, cutting analyst research time significantly.
- Engineered CAAL, an NLP relevance and classification pipeline for legal e-discovery — 10x faster document processing.
- Built custom ElasticSearch plugins that cut retrieval latency 90% for teams reviewing millions of documents.
- Designed a reusable plugin framework that halved future build time for the engineering team.
Earlier: Java Developer, P Cube IT Services (2018)
Selected projects
Independent research and builds
Where I test ideas that are too early — or too weird — for a client roadmap.
MSc thesis · LJMU 2025
Adaptive Personalized Reflective RAG
A LangGraph-orchestrated multi-agent system combining Adaptive-RAG, SELF-RAG, and PersonaRAG
- Query-routing engine that picks retrieval strategy by question complexity — 88.6% classification accuracy.
- Self-critiquing generation via SELF-RAG reflection tokens — 8–15% accuracy gain over standard RAG.
- Context-personalized responses via PersonaRAG, cutting latency 33% and reasoning steps 19%.
In progress
Is your RAG eval testing memory?
A follow-up to a finding from my GraphRAG benchmark
- In the pilot, a ‘give your best guess’ prompt let the model answer Wikipedia questions from memory. On questions whose context held no answer, F1 was still 0.35 to 0.53.
- With a strict prompt, the same check stayed between 0.000 and 0.033. Now I’m measuring how often the generator fills gaps from memory, which I didn’t run the first time.
Toolkit