I'm a Computer Engineering student at Savitribai Phule Pune University (CGPA: 9.136/10, 2023–2027) and a Generative AI Engineer who builds production-grade AI systems — from multi-agent pipelines and causal inference engines to full-stack, containerized, deployable products.
Most student AI projects are tutorials with a frontend slapped on. Mine aren't. I've shipped a 5-agent medical billing auditor that generates IRDAI-compliant appeal letters in under 15 seconds, a causal policy simulation engine over 700K+ records, and an NLI-based hallucination detector benchmarked across 5+ LLM variants — the kind of work that holds up under scrutiny from engineers, not just professors.
What I'm building right now:
- 🤖 Multi-agent RAG systems with LangGraph — document ingestion, regulatory reasoning, QA judges
- 🧠 Causal representation learning & counterfactual policy simulation under distribution shift
- 🔍 LLM factuality pipelines — NLI-based hallucination detection with probabilistic calibration
- 🏗️ Full-stack AI products: Spring Boot + FastAPI backends, Next.js frontends, Dockerized deployments
I'm actively looking for: AI/ML internships, SWE/MLE roles, and open-source collaborations where I can work on problems that matter.
I work across the full depth of an AI system — from mathematical foundations to shipped, containerized products.
Depth: Transformer architectures (self-attention from scratch), LoRA/adapter fine-tuning, DeBERTa/RoBERTa NLI-based entailment & calibration (ECE, precision-recall), counterfactual simulation under do(·) interventions, DAG identifiability (backdoor/frontdoor), multi-agent orchestration with LangGraph, RAG with hybrid sparse-dense retrieval (FAISS + BM25), Stable Diffusion, cGAN (Pix2Pix U-Net + PatchGAN), Neural Style Transfer, NLP-to-SQL, time-series forecasting, Isolation Forest anomaly detection.
Depth: Modular microservice design (Spring Boot + FastAPI tri-service architectures), async ingestion/query decoupling, SSE streaming for live agent output, JWT auth, hot-swappable model backends, containerized deployments with Docker Compose + Nginx + SSL, scalable inference serving, embedding pipeline architecture.
Depth: Vector similarity search (pgvector + FAISS), hybrid sparse-dense retrieval, relational schema design, query optimization, geospatial indexing with PostGIS, sub-second semantic lookup over 3.5M+ records, Redis Streams for real-time event-driven pipelines.
Depth: Next.js 15 App Router with SSE streaming, Zustand state management, D3.js + Recharts + Plotly for multi-layer data visualization, Mapbox geospatial maps, Framer Motion animations, Gradio/Streamlit for rapid AI prototyping interfaces.
Full-stack AI systems built with production discipline — not just notebooks.
The Problem: Hospital bills are opaque, overcharged, and nearly impossible for patients to dispute. Insurance appeal letters require deep regulatory knowledge (IRDAI/CGHS) that no patient has — and no existing tool provides.
The Approach:
- Architected a 5-agent sequential pipeline: Document Auditor → Clinical Reviewer → Regulatory Advisor → Appeal Drafter → QA Judge — each agent with a distinct role, handoff protocol, and failure mode
- Parsed hospital bills using LayoutLMv3 + EasyOCR; benchmarked extracted charges against official CGHS rates for line-item anomaly detection
- Implemented RAG over IRDAI circulars using FAISS + sentence-transformers for zero-hallucination regulatory citations
- Streamed live agent output via SSE (Server-Sent Events) for real-time UX — no polling, no waiting
- Generated IRDAI-compliant appeal letters in under 15 seconds end-to-end
Architecture:
Hospital Bill (PDF) → LayoutLMv3 + EasyOCR → Document Auditor Agent
↓
Clinical Reviewer Agent
↓
RAG over IRDAI Circulars (FAISS + sentence-transformers)
↓
Regulatory Advisor Agent
↓
Appeal Drafter Agent
↓
QA Judge Agent → SSE Stream → Next.js 15 UI
Results: Full appeal letter generated in <15 seconds; zero hallucination on regulatory citations via evidence-grounded RAG; live streamed agent output via SSE.
Next.js 15 FastAPI PostgreSQL pgvector FAISS LangGraph LayoutLMv3 EasyOCR sentence-transformers Docker
The Problem: Climate policy decisions in India are made without rigorous counterfactual modeling — policymakers can't answer "what would Delhi's AQI look like if the Clean Fuel Subsidy hadn't been passed?" Correlation-based ML offers no causal guarantees.
The Approach:
- Built a 4-module platform for Indian policymakers: real-time AQI + weather maps (IMD/CPCB data via Mapbox + Recharts), DoWhy-inspired causal engine, RAG pipeline, and live sentiment analysis
- Specified structural equation models (SEMs) capturing causal relationships between policy interventions, emissions, economic output, and public health outcomes
- Implemented ATE and CATE estimators with
do(X = x)intervention semantics; conducted sensitivity analysis (Rosenbaum bounds, E-values) to quantify robustness under latent confounding - Built RAG pipeline over MoEFCC and UN reports using pgvector for state-level evidence-grounded policy recommendations
- Implemented live sentiment analysis (BERT/RoBERTa) over Twitter + news streams with state-wise topic clustering
- Async ingestion pipelines over 700,000+ heterogeneous policy and climate records with sub-second semantic retrieval
Architecture:
IMD/CPCB/Twitter Streams → Async Ingestion → PostgreSQL + pgvector + MongoDB
↓
SCM / DAG Construction (DoWhy) ←→ RAG over MoEFCC Reports
↓
Counterfactual Simulation Engine (do-calculus)
Sentiment Analysis (BERT/RoBERTa) — state-wise clustering
↓
FastAPI Backend → React + Mapbox + Recharts + Plotly Dashboard
Results: Reproducible counterfactual policy simulations with calibrated uncertainty; state-level sentiment clustering over live news and social streams; full offline demo fallback.
Python DoWhy FastAPI pgvector React Mapbox Recharts Plotly BERT RoBERTa PostgreSQL MongoDB Docker
The Problem: LLMs hallucinate confidently. Existing detection methods are either too shallow (keyword matching) or too slow (full re-generation). There's no lightweight, calibrated, production-deployable solution.
The Approach:
- Formulated hallucination detection as a conditional inference problem: given retrieved evidence
Eand generated claimC, estimateP(entailment | E, C) - Fine-tuned DeBERTa-based NLI classifiers on domain-adapted QA corpora; evaluated calibration rigorously via precision–recall curves, ECE (Expected Calibration Error), and confidence distribution analysis across answer confidence bins
- Identified systematic degradation under retrieval noise, semantic drift, and distribution shift; exposed failure modes including overconfident contradiction misclassification and hallucination in low-evidence contexts
- Designed independent microservices for retrieval, entailment, and confidence scoring — enabling hot-swappable model backends
- Explored continual adaptation mechanisms to mitigate model drift as LLMs evolve
Architecture:
LLM Response → Dense Vector Retrieval (pgvector)
↓
Evidence Ranking & Context Assembly
↓
DeBERTa NLI Entailment Scoring (ECE-calibrated)
↓
Confidence Calibration → SHAP XAI Verdict (Gradio UI)
Results: >25% improvement in factual reliability on benchmark datasets; benchmarked across 5+ LLM variants on factuality, precision, and ECE calibration metrics.
DeBERTa RoBERTa HuggingFace RAG pgvector FastAPI Gradio SHAP PostgreSQL Continual Learning
|
Schema-aware NLP-to-SQL engine over 3.5M+ ARGO float oceanographic records. Integrates geospatial query execution (PostGIS), dynamic parameter binding, and live Plotly visualizations — enabling non-technical researchers to explore marine datasets through natural language alone. Decoupled API layer separates NL interpretation from SQL execution.
|
Tri-service platform ingesting Indian bank statements from 5 major banks, classifying transactions at 91% accuracy (sentence-transformers + FAISS), and computing a 5-dimension Financial Risk Index (liquidity, debt burden, volatility, savings stability, investment diversification). Scenario engine lets users adjust parameters and see 6–12 month cash flow forecasts.
|
|
Event-driven multi-agent system where financial market agents propose actions, negotiate conflicts, and self-enforce governance rules using PPO/SAC from first principles. Models Nash bargaining, meritocratic voting, and emergent equilibrium stability under real market data.
|
Low-power streaming 1D CNN accelerator for real-time ECG arrhythmia detection on FPGA hardware. Hardware-accelerated deep learning for medical AI — built for the FPGA Hackathon 2026. Optimized for minimal latency and power budget on embedded inference.
|
|
|
|
|
-
Hackathon Director (College-Wide) — Sole organizer and SPOC for an institution-wide hackathon spanning 160 teams and 500+ participants; owned all logistics, scheduling, judging coordination, and real-time decision-making end-to-end. Largest technical event run by a single student at the college.
-
Vice Chair, ACM Student Chapter — Directed 6+ technical workshops on ML systems, LLMs, and competitive programming reaching 1,000+ students; represented the chapter at 10+ external hackathons with consistent top-5/10 finishes; coordinated inter-college hackathon partnerships and peer research initiatives.
-
Sponsorship Lead, MPulse Technical Fest — Closed 30+ sponsors and raised ₹1 lakh+ through end-to-end acquisition (cold outreach, pitch decks, negotiations); largely funded the college technical fest with 500+ attendees.
-
AI & Computer Vision Lead, Team Vulcans Robotics — Designed the computer vision pipeline for ABU Robocon 2026 from scratch; deployed real-time object detection and localization systems on embedded hardware; mentored 4 junior contributors on model integration, quantization, deployment workflows, and code review; represented the team at state-level robotics competitions.
-
Team Lead, SIH · SKNCOE Fusion 2025 · PCCOE IGC Hackathons — Directed multi-disciplinary teams delivering full-stack AI prototypes under tight deadlines; consistent top finishes across national and college-level competitions.
Open to AI/ML internships, SWE/MLE roles, and research collaborations.
I build at the intersection of causal AI, LLM reliability, and multi-agent systems —
if you're working on something in that space, or just want to talk about a hard problem, reach out.