B.Tech Information Technology · Banasthali Vidyapith (2024-28) · GPA: 8.9/10.0

I build AI agents for domains where being wrong isn't an option: law, healthcare, finance.

My work sits at the intersection of multi-agent orchestration and safe AI deployment - designing systems with hallucination-constrained generation, citation-gated outputs, calibrated refusal mechanisms, and sentence-level audit traceability. Every agent I build knows when to answer and, more importantly, when not to.

Most LLM applications bolt a prompt onto an API and ship. The hard problems I care about are different:

• Hallucination mitigation in high-stakes domains - How do you guarantee every legal citation maps to a real statute? Every diagnostic claim traces to a medical reference?
• Adaptive agent routing - Why should a simple query and a complex one take the same path through your system? I build state machines that route based on what they discover, not fixed pipelines.
• Calibrated refusal - An agent that says "I don't know" with the right threshold is more valuable than one that always answers. I implement explicit abstention behavior when retrieval confidence falls below threshold.
• Provenance and auditability - If your system generates a recommendation, you should be able to trace every claim back through the retrieval → grounding → generation chain to its source.

LangGraph · LangChain · ChromaDB · OpenAI GPT-4o-mini · FastAPI · Pydantic · Streamlit

• Engineered a LangGraph state-machine pipeline (4 nodes, 12-field TypedDict state) that transforms raw transaction monitoring alerts into audit-ready Suspicious Activity Reports
• Designed a 3-layer triage cascade - rule-based filtering → behavioral anomaly scoring across 6 dimensions → LLM judge restricted to borderline cases only - eliminating 85-90% of false positives
• Built a citation-constrained RAG engine over a 4-document FinCEN regulatory corpus (~45 chunks) with sentence-level provenance tracking
• Implemented 11-check 5W+How validation across three severity tiers (Critical/Major/Minor) with full audit trail compilation
• Achieved 100% classification accuracy across 5 synthetic scenarios (~3,200 LOC, 15+ Pydantic schemas)

TMS Alert → [Ingest & Enrich] → [Triage & Classify] → Conditional Router
                                                          ├── ~70-75% → FALSE_POSITIVE (auto-closed)
                                                          ├── ~15-20% → NEEDS_REVIEW (human queue)
                                                          └── ~10-15% → [RAG + Generate] → [Validate & Package] → SAR Output

FastAPI · Qdrant · OpenAI Agents · RAG · Docker · fpdf2

• Architected a 5-agent orchestration pipeline (Intake → Router → Retrieval → Answer → Reporter) with adaptive reasoning paths across Indian legal codes (BNS, CrPC, IPC)
• Router Agent implements a query complexity classifier (Layman vs. Paralegal) selecting divergent reasoning paths based on input complexity
• Designed a hallucination-constrained retrieval workflow - every legal claim grounded via Qdrant vector search with statute-aware re-ranking
• Implemented calibrated refusal - agent explicitly declines when retrieval confidence falls below threshold instead of hallucinating
• Reporter Agent generates structured PDF advisories with traceable citation chains: claim → statute → provision → remedy

📄 Paper: "Legal-MVP: A Multi-Agent Agentic Framework for Citation-Constrained Legal Advisory in the Indian Judiciary" - in preparation (2026).

Azure OpenAI (GPT-4o Vision) · FastAPI · Azure Web Apps · RAG · PDF Generation

• Built a multi-modal vision-language pipeline processing ultrasound (PCOS) and X-ray (pneumonia, breast cancer) inputs through GPT-4o Vision with domain-adapted prompt templates
• Engineered confidence-calibrated diagnostics by cross-referencing LLM outputs against medical reference datasets
• Generates radiologist-style PDF reports with severity classification (Normal/Mild/Moderate/Severe) and structured finding extraction
• Designed as a unified health identity layer aligned with India's ABHA infrastructure, addressing record fragmentation across hospitals
• Production-deployed on Microsoft Azure Web Apps

🔗 Live: vitalscan-med-ai.azurewebsites.net

LangChain · FAISS · OpenAI Embeddings · FastAPI · Multi-Agent RAG

• Designed a 7-stage sequential pipeline with typed inter-stage contracts (structured JSON handoffs) for insurance claim adjudication
• Strict schema enforcement at stage boundaries achieves sub-2% hallucination through citation-gated generation
• Every output requires exact policy clause references, section identifiers, and coverage conditions
• Implemented per-decision confidence calibration enabling threshold-based routing to human reviewers for borderline cases

LLM Tooling · SQL Analytics · RAG · SQLite/PostgreSQL

• Built an intent-conditioned routing architecture dispatching to specialized backends: text-to-SQL for quantitative queries, lightweight RAG for contextual/policy questions
• Implemented bilingual (English/Hindi) natural-language support over national water datasets (CGWB, state boards)
• Response synthesizer produces citation-aware outputs with interactive visualizations

🔗 Live: neer-setu-copilot.streamlit.app

Orchestration       LangChain · LangGraph · LangSmith · OpenAI Agents SDK
LLMs & Vision       GPT-4o · GPT-4o-mini · GPT-4o Vision · Ollama (local)
Vector Databases    Qdrant · ChromaDB · FAISS
Retrieval           RAG · Semantic Search · Re-ranking · Embeddings · Text-to-SQL
Backend             FastAPI · Pydantic v2 · RESTful APIs · Streamlit
Cloud & DevOps      Microsoft Azure (Web Apps, OpenAI Service) · Docker · Git
Languages           Python · C++ · C · SQL · Bash
Safety & Eval       Hallucination Detection · Citation Verification · Provenance Tracking
                    Confidence Calibration · Ablation Studies · Error Taxonomy

Every project isn't just a one-off - they share architectural principles I've developed through iteration:

• 📚 Behavioral psychology & decision science
• 🎾 Competitive tennis
• 🌍 Global politics & geopolitics
• 💃 Dance as structured creativity
• 🏃‍♀️ Long-distance running