Abhimanyu Sikarwar

• Designed and deployed an agentic AI reporting assistant using LangChain and LangGraph; the agent runs RAG pipelines with vector search over company invoices, reimbursements, and transaction records to answer financial queries in natural language chat, replacing manual report generation.
• Built multi-agent orchestration workflows with tool-use and memory for complex multi-step financial reasoning; applied prompt engineering techniques to improve response accuracy and reduce hallucinations in enterprise financial context.
• Architected procure-to-pay platform with role-based approval workflows (Form.IO), covering PO creation, vendor selection, invoice generation, and payment via credit cards and virtual accounts; delivered in 6 months against a 12-month estimate.
• Led corporate credit card onboarding platform that evaluated creditworthiness using financial data and director scores; replaced spreadsheet ops and cut onboarding time by 50%.
• Raised test coverage from 15% to 80% and reduced production bug rate by 60% by establishing code review standards and CI/CD pipelines across a team of 5 engineers.

• Collaborated with ML engineers to fine-tune and productionize an AI underwriting engine (AUSIS) that scored insurance applications using multi-modal inputs — medical records, financial history, location-based fraud signals, and wearable health data; built Angular UI surfacing model risk scores and recommendations, cutting processing time by 40% per application.
• Contributed to production insurance fraud detection model integrating behavioral signals, claim patterns, and historical data; built real-time review dashboard visualizing model outputs, fraud scores, and detection reasoning for claims teams.
• Integrated computer-vision AI model for vehicle damage assessment; model analyzed uploaded claim videos, detected damaged parts, and predicted repair and replacement costs including labor; enabled faster claims estimates and fraud detection for duplicate or inflated claims.
• Worked with Python-based ML APIs to surface model outputs in Angular UIs; participated in model evaluation cycles and fine-tuning feedback loops; delivered enterprise-grade platform for Wipro serving 500+ users.
• Led 3 engineers to build shared component library across all AI product UIs; accelerated feature delivery by 35%.

• Cart abandonment hit 45% at checkout; integrated Stripe and Paytm with unified UX and error recovery flows; raised conversion by 25% and cut payment failures by 40%.
• EdTerra had 20% monthly churn; added blogs, quizzes, and geolocation content features; lifted engagement 40% and reduced churn to 12%.
• Site load times reached 3 seconds under traffic spikes; applied Redis caching, query tuning, and CDN; cut load time to under 1.5 seconds and tripled capacity.
• Delivered Houzzcart e-commerce platform in 4 months using Angular 12 and Node.js; launched on time with zero critical bugs.

TypeScript · React Native · Expo · Node.js · MongoDB · MCP · Cloudflare Workers

• Built an AI orchestration platform with custom and remote MCP servers, skills, projects, and patient history memory; shipped an Android app on Google Play and a web app in six weeks.
• Wrote clinical-mcp, a TypeScript MCP server exposing 60+ tools: ICD-10/11, LOINC, and HCPCS code lookups, RxNorm drug search, drug-drug interaction checks, and 35 deterministic clinical calculators; deployed as a remote MCP server on Cloudflare Workers, usable by any MCP client.
• Designed a document-reading pipeline where a vision model transcribes photographed lab reports, ECGs, and prescriptions for a text-only chat model, with extractions cached by content hash to avoid re-billing.
• Enforced safety rules for clinical use: dose and score arithmetic goes through deterministic calculators instead of the LLM, citations come only from doctor-provided documents, and memory sits behind consent controls with an audit trail.

MCP · LangGraph · Sarvam AI · Python · WebSocket · REST APIs · LLMs

• Built agentic voice calling pipeline using Sarvam AI text-to-voice; the agent calls leads, listens to responses, and generates follow-up questions from conversation context to assess purchase intent.
• Implemented MCP-based orchestration to coordinate the calling agent, response analysis agent, and CRM update agent; each agent hands structured context to the next without human intervention.
• Designed lead scoring model that classifies leads as hot, warm, or cold based on conversation transcript analysis using LLMs; reduced manual follow-up effort by filtering out uninterested leads before reaching the sales team.

MCP · LangChain · LangGraph · OpenAI API · Claude API · Python · REST APIs

• Designed multi-agent pipeline where a diet-planning agent generates weekly meal plans from user health preferences, then a procurement agent breaks meals into ingredient lists and places orders via grocery store APIs.
• Implemented Model Context Protocol (MCP) to enable context handoff between ChatGPT and Claude agents across planning and ordering steps without losing user preference state.
• Integrated location-aware store selection and real-time inventory checks to route orders to the nearest available store, reducing fulfillment gaps.

React · TypeScript · WebSocket

• Traders needed faster access to unusual options activity; built real-time analytics dashboard with WebSocket streaming; reduced decision-making time by 35%.
• Users requested customizable views for different trading strategies; implemented advanced filtering with saveable presets; increased daily active user engagement by 45%.
• Market data was hard to visualize at scale; designed responsive S&P 500 heatmap with sector-based rendering; improved data comprehension based on user feedback.

• Formerly the Ambedkar Institute of Advanced Communication Technologies & Research (AIACTR), now NSUT East Campus, Delhi.