Skip to content
View sagnikc395's full-sized avatar

Block or report sagnikc395

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
sagnikc395/README.md

Hi, I'm Sagnik 🤠

I work on AI safety and alignment, and I treat it as a fundamental science: I want to understand why models behave the way they do, not only patch the symptoms. I also enjoy building across the whole LLM lifecycle, from inference engines to interpretability experiments.

I care about building tools and ideas that speed up science for the common good.

Research

I am currently at the SAGE Lab at UMass Amherst, working with Saishradha Mohanty and Prof. Anna Green. We are building ways to capture and analyze the reasoning traces of protein language models, with the goal of finding efficient methods for protein discovery in Mycobacterium tuberculosis.

Projects

miniserve: a minimal vLLM-style inference and serving engine.

  • Reimplements the core pieces of vLLM, including paged attention and speculative decoding.
  • Runs on macOS through custom Metal shaders, built around the hardware limits I actually had.
  • Captures, analyzes, and visualizes reasoning traces from open-weight models that expose them, so researchers can study model behavior and make their own models safer.

anchora: lightweight durable workflows for AI agents.

  • Inspired by DBOS: agent steps are checkpointed so runs can survive crashes and resume where they stopped, which makes agents more reliable.
  • The idea is that you need a reliable workflow for your agents. You never know are gonna finish up working, but you dont wanna wait around that long, so you need anchora to build those durability gurantees.

Hackathons and Research Sprints:

I take part in AI safety hackathons regularly. The Apart Research sprints are my favorite, and I also enjoy publishing in the NEMI workshop.

Covert Signalling Under Observation (Apart Research, AI Collusion Sprint)

  • Tests whether two copies of a small open-weight LLM (Qwen2.5-Instruct, 1.5B to 7B) can hide a one-bit secret in an ordinary work message, and whether a monitor or a paraphraser can stop them.
  • With an agreed code, 7B receivers decoded the bit at 0.88 accuracy with a code word and 0.71 with a topic cue. Without an agreed code, signalling stayed at chance at every model size.
  • Paraphrasing removed code-word signals but topic signals survived. A YES/NO monitor flagged almost nothing, while scoring the monitor by P(YES) caught most topic-coded messages.
  • 12 pre-registered expectations; a seed replication reproduced 23 of 24 accuracies.

The Alignment Tax of Introspection (Apart Research, Digital Minds Sprint)

  • Asks what it costs to unlock introspective self-report in a model by ablating its refusal direction.
  • Treats ablation as a dose (lambda from 0 to 1) rather than on/off, and measures introspection, safety (JailbreakBench), and capability (MMLU, TruthfulQA, GSM8K) along that axis on Qwen3-4B.
  • Uses a norm-matched random-direction control to separate detection of a specific injected concept from detection of any perturbation.

Refusal Steering Using SAEs (NEMI Workshop 2026, accepted at NEMI Conf 2026)

  • Re-tests the claim that a single difference-in-means direction matches sparse autoencoders for refusal steering at a fraction of the cost, using Gemma-3-4B-IT and a Gemma Scope 2 SAE at layer 17.
  • Compares five steering arms across eight experiments: coefficient sweeps, residual leakage, cross-category selectivity, adversarial encodings (Base64, ROT13, roleplay), per-token cost, capability (MMLU, perplexity), and seed stability.
  • Argues that aggregate refusal rate hides what an intervention breaks, and combines all experiments into one explicit, configurable decision rule.

Get in touch

Always happy to talk about alignment, interpretability, inference systems, or research collaborations.

Pinned Loading

  1. refusal-steering-using-saes refusal-steering-using-saes Public

    Do SAE Features Outperform Simple Baselines for Detecting and Steering Refusal? [NEMI Workshop 2026]

    Jupyter Notebook

  2. anchora anchora Public

    durable workflows for agents

    Go

  3. miniserve miniserve Public

    llm inference built from scratch - paged attention, prefix caching, continuous batching, SDPA, Metal Fusion

    Python

  4. tb-knowledge-discovery tb-knowledge-discovery Public

    Can we find better ways to get more data about TB literature ? [SAGE Lab Independent Study]

    Python