▮ ABHINAV.APP

AI Engineer · Boston, MA

Abhinav
Karthik

I build agents, optimizers and evolutionary systems that actually ship. Currently on AI legal intelligence at State Street, and evolutionary optimisation research at Northeastern.

See the work Resume
SCROLL⌘K

00 Now

Updated Sep 2026

  • Working

    AI legal intelligence at State Street on the Evisort platform, through Xoriant.

  • Researching

    Chasing the Battleship shot-count record with exact posterior rollouts. Current dev mean: 46.35 shots.

  • Reviewing

    Two papers under double-blind review: a GA-tuned ensemble agent and adaptive GGA→SSGA switching.

  • Building

    7D Connect, a school-bus operations platform with native driver apps, and the MaineHealth scheduler.

01Selected work

Things you can
actually poke at.

Research that turned into products, products that turned into research. Four of them run live in your browser.

Also built

  • Misalignment sandbox — A simulated pharma company for studying LLM misalignment with chain-of-thought monitoring and time control.
    2025repo
  • Cognitive & emotional state tracker — Real-time multimodal tracking of 28 emotional and cognitive states from video and audio, with an emotion-aware chat.
    2025repo
  • Satellite lawn measurement — Draw a polygon on a map, get AI-segmented lawn area by zone with a confidence grade.
    2025repo
  • Sequence research lab — Local-first research dashboard for two-player Sequence with an RL training loop.
    2026repo
  • Agent long-term memory — Persistent memory for a commercial agent, plus MCP omni-channel SMS and WhatsApp.
    2025

02Research

Evolutionary optimisation,
with receipts.

Two papers with Prof. Jonathan Mwaura at Northeastern's Roux Institute, both under double-blind review. Titles are withheld until decisions land.

under review2026

GA-tuned ensemble agent for Battleship

Four Battleship strategies, one heatmap each, blended by weights a genetic algorithm learned. Averages 40.5 moves to win and beats the 2011 benchmark.

40.5
avg moves to win
42
previous benchmark (Berry, 2011)
−4
moves vs uniform-weight ensemble
4
strategies blended
Read the abstract Live demo
under review2026

Adaptive generational ↔ steady-state switching for GAs

A genetic algorithm that starts generational, watches its own elite centroid, and switches itself to steady-state the moment exploration stalls. No hand-picked switch point.

60 / 60
runs where the trigger fired
20 / 20
successes on every benchmark
0.9986
mean accuracy on Wine
2% · 10%
centroid & variance thresholds
Read the abstract Live demo

In progress

Updated 2026-07

Battleship autoresearch

An isolated record-attempt lab: minimise unconditional shots to sink a hidden fleet, with frozen holdouts and predeclared promotion gates.

Cross-fitted full-horizon rollouts averaged 46.35 shots vs 47.18 for the exact posterior marginal across 40 development boards. Not yet promoted: the interval crosses zero.

03Experience

Where the work
happened.

  1. Jun 2026 — Now

    Boston, MA

    State Street

    Now

    AI Engineer · via Xoriant

    AI legal intelligence on the Evisort contract-intelligence platform for State Street's legal organization.

    • Building LLM-driven contract analysis and extraction workflows on Evisort for a global custodian bank.
    • Retrieval, extraction and evaluation work so legal teams can trust what the model says about a document.
    EvisortLLMsRAGPythonTypeScript

  2. Jan 2026 — Now

    Portland, ME

    MaineHealth · Roux Institute

    AI Research Engineer

    Physician scheduling for MaineHealth's cardiovascular service line, with Dr. Edward Sze and Prof. Jonathan Mwaura.

    • Built a mixed-integer linear programming system that generates annual schedules for ~60 physicians across 5 subspecialty groups and 46 services.
    • Designed a cascading multi-stage solver on Google OR-Tools CBC with a registry-based constraint engine supporting 14 configurable hard and soft constraint types.
    • Block variable compression, per-group skeleton decomposition and upstream-aware constraint propagation: 91% happiness score across 8,500+ weekly assignments in under 4 minutes.
    • Shipped it as a Flask web app with drag-and-drop pre-filled uploads, live solve progress and Excel export for clinical deployment.
    PythonOR-ToolsMILPFlaskPandasopenpyxl

  3. Mar 2026 — Now

    Boston, MA

    7D Connect · Northeastern

    Lead Engineer

    Student transportation platform for school districts: web admin, REST backend and native iOS and Android driver apps.

    • Architected a Next.js 16 monorepo where one web app serves the admin UI and the mobile REST API over a single Prisma/PostgreSQL schema.
    • Native SwiftUI and Kotlin/Compose driver apps: route start/finish, GPS-stamped attendance, inspection checklists, incident reports and mileage.
    • Role model spanning super admin, vendor admin, district and driver, with vendor impersonation for support and nightly backups.
    Next.jsReactPrismaPostgreSQLSwiftUIKotlinDocker

  4. Jun 2025 — Sep 2025

    Boston, MA · Hybrid

    EasyBee AI

    AI Engineer

    Agent platform sold to mid-market businesses. Memory, latency, RAG and omni-channel.

    • Architected a persistent long-term memory system for the primary commercial agent, lifting retention by 15%.
    • Cut agent latency by ~2 seconds through empirical evaluation and context engineering.
    • Led an agentic coding team that delivered an admin dashboard with deep-research capabilities in under two weeks, plus an automated RAG knowledge-base pipeline with industry-specific chunking metadata.
    • Built an MCP-based extension for omni-channel SMS and WhatsApp messaging via Twilio while preserving context and memory.
    • Drove an AI-first initiative automating repeatable work across business and technical teams, averaging a 3.5× productivity gain.
    LLMsRAGPineconeMCPTwilioCursorPythonTypeScript

  5. Jan 2021 — May 2024

    Bangalore, India

    BucketListEr

    Founder

    A social mobile app for creating and pursuing bucket lists with people and partner organisations.

    • Conceptualised, built and released the app; connected users with like-minded people and partners offering discounts.
    • Hired and ran two intern teams (investor search, marketing) at near-zero cost.
    MobileProductFundraising

  6. Jun 2024 — Jul 2024

    India

    SFJ Business Solutions

    IT Intern

    Performance-metrics tooling for a software testing team.

    • Wrote the sections of a Python, Pandas and MATLAB testing tool that measured algorithmic speed and surfaced sub-optimal paths.
    PythonPandasMATLAB

Education

Northeastern University

M.S. Artificial Intelligence

Sep 2024 — May 2026 · GPA 4.0 / 4.0

University of Massachusetts Amherst

B.S. Computer Engineering

Jun 2020 — May 2024 · Phi Delta Theta, founding member & VP

Certifications

  • Generative AI with Large Language Models

    DeepLearning.AI × AWS · 2024

  • Generative AI: Prompt Engineering Basics

    IBM · 2024

  • Python for Data Science, AI & Development

    IBM · 2024

04Skills

The stack,
grouped by what it's for.

drag to orbit · hover for names · click a node to isolate its group

05 About

Abhinav Karthik

I'm an AI engineer who likes problems with a scoreboard. Fewest shots to sink a fleet. Fewest unhappy physicians on a schedule. Fewest seconds before an agent answers.

Master's in AI from Northeastern (4.0), computer engineering from UMass Amherst. I've shipped a long-term memory system for a commercial agent, an optimizer that schedules a cardiovascular service line, and a Battleship engine that beats a record that stood since 2011.

Right now I work on AI legal intelligence at State Street, and I keep doing research on evolutionary optimization with Prof. Jonathan Mwaura at Northeastern's Roux Institute.