Skip to content

Kameswara Sai Srikar MandaM.S. Data Science, Northeastern · graduating Dec 2026

I ship ML and LLM systems to real users, and measure where they break.

I'm Srikar, an M.S. Data Science student and teaching assistant at Northeastern. At Launch Plan, until July 2026, I built for the people using a K-12 financial-literacy app through its pilot: a dashboard for school counselors, a college-fair kiosk for students, and LangGraph agents and a RAG search for the teaching team. Before that I led full-stack work on an AI-tool marketplace used by more than 10,000 people a month.

Open to forward-deployed engineer, data scientist and software engineer roles.

Bar chart of how often three open models repeat a planted false claim or obey a planted instruction once it is retrieved.
One planted passage in 10,000 flipped 21–42% of the answers it targeted.Read the case study →

What the studies found

Each study asks one question, measures it on real data or real models, and keeps the results that came out wrong.

Measured, not argued

Most claims about how AI systems break are a screenshot of one bad answer. These put a number on one failure at a time, with the same rules each time: one factor at a time, a noise floor from a repeat run, and predictions written down before the full run.

The series: Measuring LLM Systems

How often does checking early find a winner that isn't there?

Both versions of the test are identical, so every winner is a false one. It should be 5%.

Standard test, checked each time18.7%
A test built for peeking (mSPRT)0.23%

20,000 simulated tests at real headline-test sizes · how it was measured

Three to start with

A product people used, a tool that agents call, and the tool I use to tell whether a change really helped.

  1. Launch Plan

    Real users · January – July 2026

    High-school students pick colleges and careers without seeing what those choices will cost and earn.

    The school side was ready for a pilot, with a runbook for school admins and counselors, and 25 security and correctness findings closed, 2 of them critical.

    audit findings closed
    25audit findings closed
    commits, January to July
    ~130commits, January to July
    Diagram: a student app, a counselor dashboard and a Reality Check kiosk all talk to one Node/Express API, which uses PostgreSQL and a Python financial engine. Tags mark the dashboard, kiosk and messaging work, and the security hardening of the app and engine.
  2. bostonmove

    A tool agents call · October 2026

    Ask a coding agent when the next Red Line train is and it guesses. The real answer is public, but it's split across the MBTA's API, the Bluebikes feed and a geocoder, each with its own IDs, limits and licence terms.

    On PyPI and the official MCP registry. A version tag publishes to both, with no stored token or login. 36 offline tests replay 57 recorded real responses with the clock frozen.

    MBTA requests per trip plan
    4–16MBTA requests per trip plan
    recorded real responses the tests replay
    57recorded real responses the tests replay
    A terminal running bostonmove plan from Northeastern to Harvard. It lists five options with total minutes: the Orange Line changing to the Red Line at Downtown Crossing in 25 minutes, Bluebikes in 27, the Green Line E changing to the Red Line at Park Street in 30 with a station-closure alert, the 1 bus in 36, and walking in 69, each with an arrival time.
  3. evalpair

    Did the change help? · November 2026

    Most eval tools print two scores side by side, and a 2-point gap on 160 questions looks like a result when it usually isn't. Across my studies, the same mistake kept costing time: reading a difference without an interval, or without knowing how much the model moves on its own.

    On PyPI and the official MCP registry. Its first real run reproduced the quantization study's scores exactly, and its checkers re-scored all 9,870 of that study's replies the same way. That run also caught a parsing bug two of my studies shared.

    study replies re-scored, every one the same
    9,870study replies re-scored, every one the same
    items in the three bundled suites
    470items in the three bundled suites
    A terminal running evalpair compare on two runs of Llama 3.1 8B over 160 GSM8K problems. Run A, the 4-bit build, scores 84.4% and run B, the 2-bit build, 60.6%. The difference is −23.8 points with an interval from −31.9 to −16.2, answers changed on 40.6% of items (43 worse, 5 better), the noise floor is 0.0%, and a sentence concludes the change is beyond noise. Below it, five items that got worse and five that got better, with both answers.

Experience

Full history
  1. September 2026 - Present

    Teaching Assistant, DS 3500: Advanced Programming with Data · Northeastern University

    • Teaching assistant for DS 3500, Advanced Programming with Data, a class of 108 students, working with Professor Deahan Yu.
    • The course covers data pipelines, modelling, caching with Redis and building apps in Streamlit.
    • Hold office hours and run checkoff meetings, where each student walks me through a small project so I can check they understand how it works; I also help with the class quizzes.
    • Helped the class through Git for the first time, from installing and authenticating to their first merge conflicts.
  2. January 2026 - July 2026

    Data Scientist · Launch Plan

    • Built the school side of a K-12 financial-literacy platform whose student app went into a TestFlight pilot: a counselor dashboard from an empty Next.js shell, with cohort analytics behind a privacy floor and FERPA-gated messaging to students over push and email, and Reality Check, a college-fair kiosk that shows students what a career and college choice really cost.
    • Built LangGraph workflows and MCP server integrations that let LLM agents write and tailor lessons with internal tools and outside APIs, taking content formatting and assessment assembly off the teaching team.
    • Replaced a keyword search educators didn't trust with a RAG pipeline on Pinecone and custom embedding models, fed by a data layer built with FastAPI, PostgreSQL and Airflow that pulls in content from several school districts.
    • Led a security and correctness sweep before the pilot that closed 25 audit findings, 2 of them critical, as part of about 130 commits across the mobile app, API and data layer.
    Read the case study →
  3. January 2026

    Quantum Winter School (Mentee) · Massachusetts Institute of Technology

    • A four-day program at MIT with about 50 participants, taught by instructors from MIT Physics, MIT CSAIL and Harvard SEAS.
    • Went from qubits, quantum gates and entanglement to Bell's theorem, the CHSH inequality and algorithms such as Shor's.
    • Built and ran quantum circuits in hands-on Qiskit labs and on the qBraid platform.

Latest writing

All writing

Get in touch

Questions about my work, a role, or a project idea. Send a message here or email me directly.

Srikarmks@gmail.com