Eshaan Prasad Mathakari
AI Engineer · Data Scientist
eshaan.mathakari@gmail.com · New York City · GitHub · LinkedIn
This page prints cleanly. Use your browser's print or save-as-PDF.
Summary
AI engineer who builds agentic AI systems with several cooperating agents and human-in-the-loop approval, grounded in deterministic rules, citations and audit trails for regulated industries such as banking, compliance and legal. Works on AI safety and evaluation, including red-teaming, guardrails and benchmark-based evaluation, and delivers full stack across React/TypeScript, Python and Node, and AWS.
Experience
AI Engineer · EXL
2025 – presentPitch Compass · A prototype AI copilot that helps bank sales advisors understand business clients and choose next steps.Oct 2026
- Built a prototype AI sales copilot for treasury-banking advisors that sorts business clients into archetypes, guides discovery conversations, and returns ranked product recommendations and follow-up actions.
- Designed a live agent-step panel, pipeline analytics, and an admin area for rules, change requests and governance in React, TypeScript and Zustand.
Fine Print Hawk · A workspace that checks loan covenants against their source documents and flags every mismatch.Oct 2026
- Built a loan-covenant validation workspace for commercial-lending teams that checks covenants against source documents and flags each mismatch with a link to the source page.
- Implemented maker-checker review, write-back of approved updates to a sandboxed loan system, and an audited multi-section report, with the full review journey covered by Playwright end-to-end tests.
Paper Trail · Four AI agents run business due-diligence checks, while fixed rules decide every risk score.Oct 2026
- Built an agentic know-your-business (KYB) workbench in which four AI agents (Intake, Research, Verification, Report) gather ownership, screening and adverse-news evidence, check documents against registry data, and write the due-diligence report.
- Kept every risk score, outcome and routing decision in a deterministic rules engine rather than the AI, with a full audit trail and a governance view of agent cost, corrections and guardrails.
Rulebook Radar · Eight AI agents turn a regulatory document in any language into a reviewed change record.Sep–Oct 2026
- Built a proof-of-concept regulatory change-management platform in which eight AI agents (Claude on Amazon Bedrock) translate, scope, cite and quality-check regulatory documents in any language.
- Developed a custom evaluation harness that measures accuracy against benchmark records, with low-confidence cases routed to human review and PDF, Word and CSV export after expert sign-off.
Ticket Whisperer · Six AI agents take an IT support incident from first report all the way to closure.Sep 2026
- Built a white-label agentic incident-resolution workspace in which six AI agents handle IT incidents from intake and classification through knowledge-base retrieval, response drafting, email dispatch and closure.
- Added confidence-threshold routing between automatic replies and analyst review, plus a live agent activity log, explained reasoning per agent, and analytics and governance dashboards.
Sparring Partner · An AI red-team and blue-team platform that attacks chatbots, scores the results and proposes fixes.Mar 2026 – now
- Researched adaptive, multi-turn jailbreak attacks on large language models and grew the work into a red-team/blue-team platform that tests generative-AI chatbots and agents for security and compliance weaknesses.
- Engineered the attack, judging and guardrail-fix loop with PyRIT, FastAPI and AWS (ECS Fargate, Step Functions), scoring responses against a standards-based assurance rubric and requiring human approval for every fix.
Ink Slinger · An AI assistant that redlines NDAs and purchase terms, citing a rule for every edit.Jul–Aug 2026
- Built a proof-of-concept contract-redlining assistant (Claude on Amazon Bedrock, AWS Textract) that proposes exact-text edits to NDAs and purchase terms, each citing the rule and original wording it enforces.
- Delivered a Word-style review page, native tracked-changes output with python-docx, a rule-coverage report and an optional independent verifier pass that flags missed triggers.
Sticky Fingers · A shared hub where teammates publish reusable skills that AI coding agents can find and fetch.Jun–Jul 2026
- Built an internal skill hub where teammates publish reusable AI-agent skills, served through a local Model Context Protocol (MCP) server that reuses each developer's GitHub Enterprise login.
- Wrote the skill format specification, a searchable catalog with usage and maturity signals, and one-line Windows and macOS installers that configure corporate certificates and register the hub with Claude Code.
Inbox Butler · Two automation pipelines that replaced manual inbox monitoring and data entry for order processing.2026
- Built two AI workflow-automation pipelines that replaced manual inbox monitoring, data extraction and data entry for legal and commercial order processing, covering UCC lien search results and a shared order mailbox (Microsoft Graph API).
- Routed each order to a direct database write or to headless Playwright entry in a system with no API, secured with Azure Managed Identity and a shared audit log; SME testing on live orders estimated about a 70% improvement in operational efficiency.
GenAI Intern · EXL
2025- Worked with the GenAI Center of Excellence on LLMs and AI agents for client-facing AI solutions.
Data Science Intern · Sagility Healthcare
2022 – 2023- Designed and optimized a data pipeline in Azure Synapse that brought data together from multiple sources.
- Built Power BI dashboards for stakeholders and validated data with SQL and MS Access.
Co-owner and operator · Sneaker marketplace
2021 – 2024- Co-ran an online sneaker marketplace, handling sourcing, vendor relationships and pricing.
Research and personal projects
- Antelligence · Ant-colony teamwork for swarms of tiny robots, tested in simulation and aimed at medicine.
- slaya · Small, fast models that answer yes/no and probability questions so a big AI doesn't have to.
- Swarm Oracle · Several AI agents forecast the same question, and the ones with better track records get more say.
- Honest to a Fault · Can you make an AI model more honest by turning up its internal "honesty" signal?
- ScoutDocs · Gives AI coding assistants the documentation for the exact software version you are using.
- The home agent fleet · An always-on team of AI agents that runs on a Mac Studio in my apartment.
Education
- MS in Data ScienceUniversity of Arizona · GPA 3.7/4 · 2025
- Bachelor's in Computer EngineeringSRM Institute of Science and Technology · GPA 8.9/10
Certifications and posters
- Poster, IEEE EMBS 2025UCLA · Antelligence
- Poster, SBP-BRiMS 2025Carnegie Mellon · Antelligence
- AWS Certified Generative AI Developer – ProfessionalAmazon Web Services · April 2026
- Agentic AI Engineering programADaSci · 2025
- Machine learning certificationAlteryx · June 2024
- AWS Certified Cloud PractitionerAmazon Web Services · January 2022
Skills
- Languages
- Python, TypeScript, JavaScript, SQL, Rust, Solidity, C++, R
- AI
- Claude, GPT, Llama, Qwen, Phi, LangChain, Langflow, PyTorch, Hugging Face, MCP, MLX, llama.cpp, Ollama
- Infrastructure
- AWS, Azure, Docker, launchd, Cloudflare, RunPod, WireGuard, Matrix
- Data
- PostgreSQL, MongoDB, SQLite, ETL, Power BI