AI Engineer · Computational Modeling · Interpretability

Describing the World
One Model at a Time

I ship production AI infrastructure, agentic pipelines, evaluation tools, and multi-agent simulation systems.

Get in touch Download resume GitHub
◈ Krishnasai Addala
Human Researcher
Analyze
Examine the material conditions that influence Society computationally
Then, Improve them

"Models describe the world.
Only We can change it."
CMU MSBA IIIT Delhi Human Centric Systems
● TinyNLA
Natural Language Autoencoder
Verbalize
Read a model's internal state at one word and describe it in plain English. Rebuild the state from the description alone.

"Tell me Machine, what are you Thinking?"
GPT-2 · layer 8 Interp. GRPO
○ PolicySwarm
Opinion Dynamics Simulator
Simulate
Seed Users on a synthetic social network with different opinions. Perturb the network
Then, observe how Opinions change

"Who hears the shock decides what it does."
120 agents · 30 seeds Discourse Modelling Social Computing
◆ Thought Reader
Plain-Language Model Reports
Explain
Turn a model's internal state into a short report anyone can read, with a label for how far to trust it.

"Themes, not claims."
TinyNLA FastAPI D3
0/0 1 / 4
IIIT Delhi BTech CSSS . Carnegie Mellon University MS . Ex-Capgemini NYC Researcher and Engineer

Experience

Jun – Nov 2025

Capgemini · Associate Gen AI Developer Consultant · New York, NY

Shipped LangGraph and OpenAI API tooling to financial-institution clients on GCP. Owned LLMOps across model versioning, multi-environment deployment, and Airflow-orchestrated pipelines. Designed scalable ETL and streaming systems on Pub/Sub and BigQuery.

Jan – Apr 2025

Project Geminae · ML Engineer, CMU Capstone · Pittsburgh, PA

Delivered predictive well-selection models for an energy-sector client across 30+ operational variables. Built SHAP-driven feature importance outputs formatted for engineering stakeholders.

Jun 2023 – May 2024

MIDAS Lab, IIIT Delhi · Undergraduate Research Engineer · New Delhi, IN

Built a knowledge-graph-augmented LLM system for physics question answering, contributing to three published papers. Coordinated three cross-functional ML sub-teams.

Current Work

LLM Infrastructure & Interpretability Platform

● Active

An evaluation server for open-weight LLMs that tests behavior under pressure and records how internal activations shift. The sycophancy suite is built; jailbreak, faithfulness and hallucination suites are designed.

FastAPISSEPyTorch CUDAHuggingFaceMLflow
Read all about it

AI Governance Policy Simulator

○ In revision

Agent-based simulation of AI policy discourse: 120 agents on a scale-free network, 30 seeds per condition. Tests how shocks, media and the choice of opinion-distance metric change where a population's opinions end up.

PythonNumPySciPy NetworkXspaCyABM
Read all about it Methods paper: Hidden Choices

TinyNLA — Natural Language Autoencoder

● Active

A Natural Language Autoencoder for small models: it describes GPT-2's internal state in plain English, and those descriptions rebuild the state better than summaries of the input text. Includes a Thought Reader tool and plain-language reports.

PyTorchCUDAGPT-2GRPOFastAPI
Read all about it

Skills & Projects

Fullscreen ↗
Hover over a node · Drag to rearrange · Scroll to zoom

Publications