Srushti Chaukhande

Srushti Chaukhande serious software engineer (pre‑ai era) to now a wannabe empirical researcher and a occasional storyteller

Read my chain of thought About me Projects Professional timeline
(01)

Hi, I’m Srushti

Srushti Chaukhande

I love world building and the perfect balance of edge cases and order of things hence I do what I do: deep dives on my favorite research topics (interpretability, AI control, misaligned behaviours in AI) and wonder how world order can be brought to these open problems.

I love a good whodunnit mystery which is why I was assigned a lot of memory leak SEV-2s back in the day (when an AI couldn’t parse your codebase or valgrind outputs and it was up to this sensitive naked eye to save the next release).

I am currently learning the etiquettes of empirical research and honing my writing abilities without an AI helping me paraphrase every thought in my mind.

Email GitHub LinkedIn X
(02)

Current Projects

Exp. 01shipped

One Security Scanner for All Code

How do you add a new language to Vanir without writing a parser from scratch?

Worked for Google's Android Security team, as part of my UW iSchool capstone, on Vanir: Google's open-source scanner that reads source code to find security patches that never made it in. It only understood C/C++ and Java, and every new language meant weeks of hand-writing a fragile parser from scratch.

problem
each language needed its own hand-built grammar: weeks of work, unreliable community versions, nothing reusable for the next one
solution
I plugged in tree-sitter, the parser behind GitHub and Neovim, through one shared layer. Python is now a pip install; any new language is three short methods
proof
caught a critical (CVSS 9.0) unpatched vulnerability in a real Python project, and parsed Django, Ansible and Scrapy, 5,000+ files, with zero errors

Project deck ↗google/vanir ↗Design doc ↗

Exp. 02in progress

SpecTrace

Can a model tell when code does everything right but says the wrong thing?

An AI agent that checks cryptographic code against the RFCs it claims to implement, hunting bugs fuzzers walk past. 33 judgments on OpenSSL's TLS 1.3 so far.

Code ↗Write-up →

Exp. 03done

What Refusal Looks Like Inside Qwen

When Qwen refuses very different harmful requests, is it thinking the same thing each time?

I ran Anthropic's Jacobian lens over Qwen3.6-27B at the moment it refuses, using Transluce's WeirdChat transcripts. Self-harm, drunk driving and lasers at aircraft all light up the same hidden words around layers 37–47 (unsafe, prohibited, and oddly, cybersecurity). A plain logit lens misses them entirely.

Code & write-up ↗Jacobian lens · Qwen3.6-27B

Exp. 04done

Why Claude Says No to Safety Research

When Claude Opus 4.5 refuses a safety-research task, is it objecting to the work, or unsure the request is legitimate?

I extended Singh's safety-refusals experiment. Adding one line saying Trust & Safety had signed off took refusals from 80% to 0% over 50 runs, with the task itself unchanged. The reasons the model gives for refusing don't predict what changes its mind.

Code & write-up ↗Claude Opus 4.5 · n=50 per condition

(03)

My Timeline

  1. degreeUniversity of Washington, Seattle

    M.S., Information Management

    GPA 4.0. Where the pivot from building systems to studying them began.

  2. Jan 2026 – nowSeattle, WA

    Technical Program Coordinator at University of Washington

    Unifying student data from 8+ sources into one advising database and reporting platform.

    What I did
    • Designing a longitudinal advising database integrating 8+ heterogeneous sources, unifying ~4,000 undergraduate records.
    • Architecting a web-based reporting platform consolidating UW enterprise student data, advising check-ins, and survey data.

  3. Apr 2023 – Jan 2024Gurgaon, India

    Technical Lead at Capgemini Engineering

    Led performance engineering at 20K-user scale, and chased down a vanishing memory leak.

    What I did
    • Led performance engineering with 5K-20K user load tests to identify 15+ scalability issues and resolve 8 critical defects through cross-functional and vendor RCA.
    • Resolved a critical Informix memory leak at 20K-user load via custom memory-pool diagnostics, improving throughput by 22% and preventing system-wide crashes.
    • Cut REST API response time by 97.5% under 20K+ concurrent requests by refactoring backend DB queries.
    • Led FIPS140-2 certification across three releases by validating cryptographic modules and execution paths across 4+ core services.

  4. Apr 2021 – Mar 2023Gurgaon, India

    Senior Software Engineer at Altran (acquired by Capgemini Engineering)

    Built Python pre/post-upgrade health checks (80% fewer failed upgrades), moved service auth from OAuth ROPC to Client Credentials on Azure AD, and terminated TLS at HAProxy for a 35% speedup.

    What I did
    • Built a Python-based pre/post-upgrade health check framework running 20+ automated validations, reducing upgrade failures by 80%.
    • Replaced OAuth ROPC with Client Credentials for Unified Messaging using Azure AD and EWS for secure service-to-service auth.
    • Offloaded TLS via HAProxy SSL termination, delivering a 35% performance improvement and simplifying secure traffic management.
    • Mitigated RTP traffic floods by implementing packet-stat based endpoint inactivity detection and timeout-driven call teardown.

  5. Apr 2019 – Mar 2021Gurgaon, India

    Software Engineer at Aricent

    Ported core C++ components to 64-bit Linux, enforced product security baselines for 40% fewer vulnerabilities, and upgraded CiscoSSL to 1.1.1.

    What I did
    • Migrated critical C++ components to 64-bit Linux and implemented product security baselines, reducing vulnerabilities by 40%.
    • Upgraded CiscoSSL to v1.1.1, reducing exploit risk by 30% and strengthening cryptographic security.

  6. degreeDr. A.P.J. Abdul Kalam Technical University, India

    B.Tech., Computer Science and Engineering

    GPA 4.0, First Division with Honors. Where it all started.

(04)

Skills

  • Python
  • C++
  • Java
  • JavaScript/Node.js
  • Distributed Systems
  • Scalable Backend Architecture
  • System Design
  • REST APIs
  • Linux
  • Docker
  • Kubernetes
  • SQL
  • NoSQL
  • Cloud (AWS/Azure)
  • CI/CD
  • Debugging & Testing
  • Agentic AI Systems
  • LLM Pipeline Design
  • Prompt Engineering
  • Research Methodology
  • Open-Source Development