Vansh Gupta

AI Engineer · Production-Grade Intelligent Systems

Building intelligent systems that move from notebook to production — RAG pipelines, LLM agents, financial analytics, and NLP applications.

I'm an AI Engineer focused on building reliable, production-grade systems that turn large language models into practical tools. My work spans the full lifecycle — from designing retrieval pipelines and agentic workflows to deploying full-stack APIs and building domain-specific NLP applications.

Right now I'm deep in earnings call intelligence and corporate compliance automation, two domains where AI can meaningfully reduce analyst workload and governance risk. I'm also actively exploring geopolitical risk modeling and domain-specific fine-tuning.

Languages

AI & ML

Backend & Databases

Frontend

Cloud & DevOps

Conversational analytics platform that turns plain-English questions into grounded, SQL-backed answers over your own data

GINA lets users upload a CSV, ask questions in natural language, and receive answers backed by real SQL execution — not model hallucination. It follows Table-Augmented Generation (TAG): the model reasons with the table, not instead of it. A live streaming pipeline (planner → SQL → execution → narration) runs transparently over SSE, so users can see exactly how their answer was produced.

Key highlights:

• PII shield runs client-side before any data leaves the browser
• Tiered SQL generation: deterministic templates → Groq → Hugging Face, with automatic fallback
• pgvector-backed semantic schema profiling for accurate query routing
• Full SQL disclosure and expandable verification queries in the UI
• Deployed and live with Supabase Auth, AWS S3, and PostgreSQL

Next.js Fastify PostgreSQL pgvector Groq Hugging Face AWS S3 SSE TypeScript

Finalist — Meta PyTorch OpenEnv Hackathon 2026 · The first open-source RL training environment for enterprise expense compliance

An OpenEnv-compliant Reinforcement Learning environment that simulates how a corporate compliance officer audits employee expense claims. The agent receives a ticket, searches a 15-rule policy document, requests missing documents if needed, and resolves each claim with Approve, Reject, or Escalate. Three difficulty tiers — easy, medium, and hard — test progressively more complex multi-turn reasoning.

Key highlights:

• Fully OpenEnv-compliant: reset, step, state, grader, and baseline endpoints all validated
• LLM agent (Llama-3.1-8B) achieves 0.90 / 0.80 / 0.70 on easy / medium / hard vs. rule-based baseline of 0.78 / 0.61 / 0.34
• Policy document is a plain policy.md — any company can swap in their own rulebook
• Grounded in Indian corporate compliance norms: ₹-denominated limits, GST, WFH, and seniority rules
• Live on Hugging Face Spaces; Dockerized for local deployment

Python FastAPI Reinforcement Learning LLM Agents Docker OpenEnv Hugging Face

Multimodal pipeline that models earnings calls as pressure-driven information environments to extract short-horizon market signals

Most earnings call analysis flattens the entire transcript into one sentiment score. This project goes further by treating the call as a pressure-sensitive interaction system — modeling the moments where a manager's wording, vocal delivery, and response behavior diverge under analyst questioning. Features like tone_divergence, specificity_under_pressure, and response_latency are designed to capture signals that markets may underreact to.

Key highlights:

• Speaker diarization with pyannote.audio; transcript alignment with WhisperX
• Structural segmentation into prepared remarks, analyst questions, and management answers — Q&A weighted 3× over scripted remarks during aggregation
• Text (FinBERT), audio (wav2vec2, openSMILE), and interaction features (text–audio divergence, hesitation under pressure)
• Leakage-aware evaluation with strict time-based train/test splits
• Targets: next-day return, 5-day return, realized volatility, earnings surprise, downside risk proxy

Python PyTorch LightGBM FinBERT wav2vec2 WhisperX pyannote.audio DuckDB Parquet

Graph Neural Network framework for detecting multiple concurrent cardiac conditions from 12-lead ECG signals

A medical ML project that reframes ECG analysis as a graph problem. Rather than treating each lead independently, it constructs graph representations of 12-lead signals that capture spatial-temporal correlations — then applies GNN message passing to predict multiple simultaneous cardiac pathologies on the PTB-XL dataset.

Key highlights:

• ECG signals converted to graph structures preserving inter-lead spatial relationships
• NeuroKit2 for physiological feature extraction (R-peaks, HRV, waveform morphology)
• Automated hyperparameter tuning with Optuna across GNN architecture configurations
• Multi-label output handles overlapping arrhythmia patterns common in real clinical data
• Evaluated on F1 and AUC-ROC per label — metrics that matter clinically, not just statistically

Python PyTorch Geometric NeuroKit2 Optuna Scikit-learn PTB-XL Medical ML

🔬 Currently building

• Pressure-aware multimodal earnings call intelligence (text + audio + interaction)
• OpenEnv-compliant agentic compliance systems with RL and LLM training pipelines (SFT + GRPO)

🌱 Actively exploring

• Geopolitical risk modeling using LLMs and structured data
• Domain-specific LLM fine-tuning for finance and legal
• Agentic AI patterns: planning, memory, multi-turn tool use

I'm open to collaborations on AI engineering projects, especially in financial analytics, agentic systems, or applied NLP.

• 📧 Email: vanshgupta1810@gmail.com
• 💼 LinkedIn: vanshgupta1810
• 🐙 GitHub: VanshGupta18