Kameswara Sai Srikar Manda

• Teaching assistant for DS 3500, Advanced Programming with Data, working with Professor Deahan Yu.
• Helping a large class through Git for the first time, from installing and authenticating to their first merge conflicts.

• Worked on a K-12 financial-literacy platform: a mobile app that went into a TestFlight pilot, plus a dashboard for school counselors.
• Built the counselor dashboard from scratch: cohort analytics behind a privacy floor, and FERPA-gated messaging to students over push and email.
• Built Reality Check, a kiosk app for college fairs that shows students what a career and college choice really costs.
• Kept the financial engine's numbers honest (tax brackets indexed to the projection year, state-tax data fixes) and closed 25 security audit findings, 2 of them critical.
• Built LangGraph workflows and MCP server integrations that let LLM agents write and tailor lessons using 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.
• Built a data layer with FastAPI, PostgreSQL and Airflow that pulls in content from several school districts.
• About 170 commits across the mobile app, API and data layer.

• On-site program in Cambridge, MA, working hands-on with quantum computing in Qiskit.

• Led full-stack work on an AI-tool marketplace (React, FastAPI, AWS, SQL) that grew to 10,000+ monthly users and a catalog of 1,900+ tools.
• Found where users struggled to search and reworked the query flow and indexing: search 30% faster, API latency 40% lower.
• Set up CI/CD with GitHub Actions and Postman tests, keeping the platform at 99.5% uptime across releases.

A reader that shows what a research paper stands on, what it argues with, and what it takes on faith, quoted from its own LaTeX and pinned to the page. 197 papers with 437 walkable links between them.

Re-analysing 22,743 real headline A/B tests from Upworthy, and a 64,000-customer email test, to see how much the analysis changes the answer. The typical test could only detect a 63% lift.

A command-line tool and MCP server that looks at a Git repository, says in plain English what state it's in, and gives the exact commands to fix it. 29 repository states and 33 error messages covered, 246 tests, published on PyPI.

A retrieval-poisoning study: plant a single false or malicious passage among 10,000 that a question-answering system reads from, then measure who it reaches, how often three models repeat it, how far it spreads, and what simple defences cost. Retrieval decides who is exposed: a planted passage reached its target and close neighbours, and almost nothing else.

An agent with long-term memory learns a fact about you, the fact changes, and later it's asked. The common design gave stale answers 23–43% of the time.

Take tasks with checkable answers and reword the prompt harmlessly: a synonym, bullets instead of prose, “please”, one typo, shuffled options. Answers changed on 4–7% of items per rephrasing, and 18–29% of items gave more than one answer across the prompts.

LLMs now grade other LLMs. Order decided many verdicts, padding a wrong answer helped it, stated confidence counted against an answer, and two of three judges favoured their own work.

Tool-use advice is mostly about the interface: fewer tools, clear names, good descriptions. Single calls were solved and the interface changes moved success by under 5 points.

Long-context tests usually hide one fact in filler and ask for it back. The fact alone was found 98 to 100% of the time at every length.

A racing game that runs in a browser tab: 12 worlds, 14 cars, nothing to install, with the skies, cars and sound generated in code. 33 → 85 fps on the racing frame after I measured where the time actually went, with render and bundle-size budgets in CI so it stays there.

A K-12 financial-literacy platform. The student app went into a TestFlight pilot with zero type errors (from 675), three accessibility passes behind it and 25 audit findings closed.

A companion that lives in your git repository and grows with how you commit, including the commits AI agents write for you. On PyPI the day it was started.

Predicting whether patients found a drug effective from what they wrote about it, across 366,815 reviews. Leak-free, XGBoost gets 77.6% accuracy and 0.85 ROC-AUC on 73,363 held-out reviews; the leaky version had claimed 78.6% and 0.86.

A way to make Gaussian processes cheaper by giving the important parts of the data better approximations. The published +41% gain was a units bug; corrected, it's about −9% (−9.9% and −9.2% in two runs).

A multi-object tracker that learns to match people across video frames, instead of using the usual hand-built matcher. 82.6% MOTA on MOT17 validation, at a 0.9 weight on geometry; pure learned matching scored 20.2%.

Northeastern University

Institute of Aeronautical Engineering (IARE)