Abhinav Arvind
Abhinav Arvind

Abhinav Arvind

I build multi-agent systems and AI-native applications end to end, with the tests, evaluations and guardrails that let them hold up in production. Driven by intellectual challenges and interesting problems.

UIUC '26 · Finance + Data Science · CS minor

More about me
Other interests
I'm a religious follower of F1, love playing poker, connect with nature by climbing mountains, and am exploring my creative side through a novel I'm writing.
Looking for
Forward-deployed, founding and applied AI engineering roles on teams shipping AI products to real users.
Location
SF, NYC or remote. Open to other locations depending on the role.
Sponsorship
US permanent resident, no sponsorship needed.
How I build
I design systems around clear boundaries and verification: typed interfaces, deterministic checks where correctness matters, and tests and evaluations on every change. I use AI-coding tools to build efficiently, and ensure safety by reviewing everything that ships.
With clients
Scoping problems with stakeholders and explaining tradeoffs, from client consulting work and from building solo for a staffing firm.
What drives me
Ownership, intellectual rigor and fast feedback, more than any one industry.

Work

What I've built and done

Engineering

2025 – now

  • Confidential-aware agent pipeline for recruiting

    Freelance AI Engineer, solo · Sep 2026 – present

    A multi-agent system that finds and screens candidates for a staffing firm while keeping each client's confidential project details inside the firm. The firm spends $2M+ a year on third-party search and data providers; the system is projected to save $1M+ of that a year.

    Python · FastAPI · React/TypeScript · SQLite · Claude, OpenAI and local models

  • Autonomous multi-agent investment system

    Independent · Oct 2026 – present

    An organization of AI agents that runs an investment cycle end to end, from choosing a policy to placing paper trades. Authority is enforced in code: models choose only among options the code allows, and a deterministic risk officer is the only one who can approve a trade.

    Python · pydantic · Claude and local models · paper brokerage API

  • Innova Solutions

    Data Science Intern · Atlanta · Jun – Aug 2025

    • Built a statistical validation and scoring tool for an AI-powered ETL agent, using synthetic test data and structural-sampling tests to catch transformation errors before production.
    • Built predictive models on 1M+ insurance claims to flag ones likely to stall, and presented a proof of concept that runs the model at hospital claim submission. Projected to cut resolution by about 11 days (260+ hours) per claim for 100K+ affected claims a year.
    • Built a reusable Python framework for data transformation, exploratory analysis and clustering that cut pre-analysis time by 40%+.

Trading

2024 – 2026

  • Derivatives and Trading Academy

    Student Trader · UIUC · FIN 412, 416, 428 · Aug 2024 – May 2026

    • Built Python research tooling for evaluating strategies: backtesting and analytics on large intraday datasets, options pricing and volatility models, and robustness testing with Monte Carlo simulation and bootstrap confidence intervals.
    • Built real-time dashboards and position-sizing models for the Rotman International, CME and Bloomberg trading competitions, and optimization and scenario models for IMC Prosperity 4.

  • FIN 419 Real Client Managed Portfolios

    Student Investor · UIUC · Aug – Dec 2025

    Built vector autoregression models on macroeconomic time series (ISM PMI, CPI, the yield curve) to generate growth and inflation signals for a $500K real-money portfolio managed under CFA Institute standards.

Other

2022 – 2025

  • Illinois Business Consulting

    Senior Consultant · Sep 2022 – May 2024

    Four client engagements across robotics, food distribution, generative AI and insurance. Evaluated GenAI and agentic workflow platforms for a large insurer's investment arm and presented an AI adoption roadmap to senior leadership; built market sizing for a generative AI startup and pricing and revenue models for a robotics manufacturer.

  • Kotak Mahindra Bank

    Corporate Banking Summer Intern · Mumbai · Jun – Aug 2024

    Built a turnaround plan for an overleveraged client to recover $30M in outstanding loans, and screened 20+ companies in the infrastructure engineering sector for the deal team.

  • Equity Research Association · igniteXL Ventures

    Student Analyst, Jan 2024 – Jun 2025 · VC Deal Sourcing Extern, Mar – Apr 2024

    Fundamental equity research and investment recommendations, and sourcing and screening early-stage startups for a venture capital firm.

Contact

Get in touch

Engineering · Freelance

Confidential-aware agent pipeline for recruiting

A staffing firm finds candidates by sending keyword searches to paid data providers, and recruiters filter the results by hand. The hardest part of that filtering depends on details that can't appear in a public job post or a search query: who the end client is, what the project is, and what really matters for the role. I'm building a multi-agent system that does this work while keeping those details where they belong.

How it works

What makes it trustworthy

  • Information boundaries enforced in code. Agents that touch the outside world have no code path to confidential data, and automated tests check the import graph on every change.
  • Humans at every handoff. The criteria, the public spec and the final shortlist all need human approval.
  • Full accountability. Every model call is logged with its prompt version, model, cost and the human decision that followed. Every run records exactly which configuration produced it.
  • Evaluation treated like a held-out test set. Quality is measured against the firm's past hiring outcomes, which are sealed away from the agents and from prompt tuning. Configurations are frozen before test runs, a fingerprint register stops jobs used for tuning from becoming test cases, and leakage is measured separately with planted canary secrets.
  • Model-agnostic. Each agent can run on a different provider or model, including local models for low-cost testing.

Status

  • WorkingThe internal pipeline runs end to end, and the firm has reviewed its output on live roles.
  • In progressDeployment and integration with the firm's internal systems, including its applicant tracking system; browser-based sourcing; the first measurement against historical hires.

Stack

Backend
Python, FastAPI, SQLite
Frontend
React, TypeScript
Models
Claude API, plus OpenAI and local models through one provider layer
Engineering · Independent

Autonomous multi-agent investment system

An investment-management system built as an organization of agents with explicit authority, a provider-neutral model layer, reproducible tests, and human gates on anything that can move money. Paper trading is wired; live trading stays off until it is formally enabled.

Architecture

Design decisions

  • An organization with mechanical authority. I rejected a single prompt switching personas, which leaves no auditable disagreement, and agent frameworks, which hide control flow behind abstractions. Who may write what is enforced in code, so even a misbehaving coordinator cannot trade.
  • Models choose inside options the code computes. Services compute the facts and the permitted choices; the model chooses and explains; services validate the result. No model output can authorize a trade, change the mandate or exceed a limit.
  • Swap the model, never the agent. Each agent has one implementation. Tests replace the model client with recorded replies, so end-to-end runs are reproducible and a stale recording fails loudly.
  • No single switch can go live. Secrets come only from the environment and are redacted in logs. Live trading needs three independent factors.

How I build it

I designed an agent-driven build pipeline. Tasks are queued in the repository; a scheduled cloud agent takes one at a time, implements it, runs the tests and commits to an integration branch. I review and merge each batch, and anything touching risk, the ledger, execution or secrets needs my approval.

The checks are what make this safe: about 480 offline tests in tiers (unit, contract, authority, collaboration and end to end), reproducible model-reply fixtures, and a separate evaluation harness for each model-backed agent.

Status

  • BuiltAll six roles and the coordinator, the model layer with local and paid models, market data, paper brokerage, reconciliation, kill switch, evaluation harness and logging.
  • In progressA stricter live rule set, real-time price sizing, evaluation with frontier models.
  • PlannedAlways-on scheduler and API, notifications, a web UI, a separate execution process for brokerage credentials, and a research platform for backtesting and promoting strategies.

Stack

Core
Python, pydantic, httpx; JSON and JSONL state
Models
Claude API and local models
Planned
FastAPI, Docker, cloud hosting, Postgres, React/TypeScript