About Me
- Hi! I’m Sathiyanarayanan Palani.
- I'm an Independent AI safety researcher and software engineer with 9+ years of production experience, transitioning full-time into mechanistic interpretability and LLM safety research.
- Interested in (Interpretability / AI Safety / Beneficial Deployment / Founding / Field Building)
Professional Skills
Skills
| Category | Technologies |
|---|---|
| Languages: | C, C++, Python, Shell & Batch basics |
| Operating System: | Linux, windows |
| Communication Protocols: | UART, I2C, SPI |
| Networking Protocols: | TCP/IP, UDP, HTTP, MQTT, SQL, Postgres, Elastic, Prometheus, OTLP Basics, WebRTC, Websockets |
| SCM: | |
| Tools and Libraries: | Gtest, FFF, Jenkins, Git, MySQL, OpenCV, Tkinter, paramiko, SCP, pandas, ASTREE, lark – parser, matplotlib, numpy, plotly, libdatachannel pyYaml, rich, pytest, polars, multiprocessing, elasticsearch, kibana, grafana, prometheus_api_client, diskcache, dash, sqlalchemy, apscheduler, aiortc, asyncio, websockets etc |
| Web Technologies: | |
| DevOps & Deployment: | |
| Infra as Code, K8s, Helm & GCP: | |
| Machine Learning: | |
| Mech Interp: | |
| PyPI: |
Work Experience ( 9+ years )
- ETL-Pipeline - Scalable Architecture & Scripting done for aggregating errors in timeseries metrics stored in centralized tsdb cluster to populate elastic indices.
- Telemetry-UI - Scalable Design & Scripting done for visualization of collected KPI's using improved MVC architecture & local cache.
- EBDT - Proposal, Prototype of Event Based Distributed Tracing architecture for enhanced telemetry collection process, drawing inspirations from Open Telemetry Project for better observability.
- QTOPP - Development of Queued tasks on Process Pools, as headless CICD runner, for scheduled statistical calculations to share refined results via SMTP.
- V&V POC - Evaluating, Prototyping to ensure the infrastructure for implementation of assisted verification and validation of designs by leveraging Pre-trained AI Models.
- Responsible for Python Automation test framework, modification, maintenance & Improvement. Handling Development, Testing, setting up CICD Nightly using Github actions with custom runners for Python packages, being part of automation telemetry team
- Implementing Serial-Over-Lan using IPMI tool & tee object using Powershell command, with monitoring thread for retrying connections during target reboot & Worked on POC for uploading register values to influx DB, using low-latency InfluxDB-line-protocol, for Storage, Visualization & analysis. Collecting Performance KPI's from Thermal & Power sub-system via Redfish, based on DMTF Docs, using ODATA queries that works on top of REST APIs
- Worked on problem statement definition, framing requirement specification, architecture design, development, testing, documentation & release for OSS PyPI Packages.
- The Packages have garnered around 86k+ life time downloads organically via GitHub and PyPI.
One of the first 5 employees. Stayed as the company grew to ~300 over 4.5 years. Played a critical role in shaping initial phase of prototypes and guiding peers, operating as an early-stage key contributor during the formative phase of the startup.
- Responsible for writing test scripts for automated validation of measurements in python based on HIL scenarios from DOORS.
- Responsible for HiL process Automation via SSH using python Libraries like paramiko and SCP.
- Responsible for development of Custom framework and reusable components in Python & Bash for automated test case generation using pandas data frames for pre-HIL execution task.
- Responsible for Runtime dynamic analysis of ADAS utility modules in c using ASTREE.
- Responsible for unit testing, L2 testing using Google test framework in c++ based on Scenarios (SIL).
- Responsible for Multi-Display Broadcast over local Network using TCP/IP.
- Responsible for Data Visualization of sensor plots in Aerospace Applications using python & plotly (3d scalar plots).
- Responsible for visualization of volumetric data in the form of normal distribution for the development of algorithm on event classification using Matplotlib.
- Responsible for Embedded Software unit and functional Testing.
- Responsible for Embedded Software Development of iElektron Security Products.
- Development of multi-branch CI/CD pipeline in Jenkins for Unit Testing with github app integration for events and mail, slack integration for notifications.
- Responsible for development of semi-automated hardware testing platform for Renesas Microcontrollers using Iron-python Renesas APIs
Education
- Graduated with a first class in Electrical and Electronics Engineering.
- Final Year project done using OpenCV and Raspberry Pi for Automatic Attendance System using Face Recognition.
- Completed with 85%
- Completed with 85%
Proposals
Element 14 Road Test
Proposal written on the Test strategy & Test procedure to be followed for testing Xsens MTI-680-DK GNSS/ISS Development Kit.
More >Cluster Monitoring
Proposal written on Monitoring clusters, for Visualization of node observability metrics, involved in Intense workloads like Training ML models.
More >River Banks, Proprioceptors & An Intangible Network
Transdisciplinary Movement redefining Information System (IS) integral to Socio-Ecological-Technical System (SETS)
More >Monosemanticity-et-Interpretability
Decomposing Multi-Layer Perceptron (MLP) activations using Sparse Autoencoders (SAEs) to identify monosemantic features via sparse Dictionary Learning.
More >Activation-Steering-et-Ablation
Experiments on Activation Steering & Ablation to understand & Demonstrate Controllable Mechanistic Interpretability through Real-Life Applications on Multi-Layer Perceptron (MLP).
More >Feature-Ranking-et-Throttle-Testing
Throttle-testing an MLP toy model by using an attached SAE to decompose its hidden representation into ranked sparse features, then applying surgical ablation and activation steering feature-by-feature (from least relevant to most relevant) to quantitatively & statistically measure each feature's causal impact on every output group.
More >LLM-Emotions
Probing emotional representations in Gemma 3 1B IT using a Sparse Autoencoder on Layer 13 activations — discovers 12 disjoint emotion clusters, recovers Russell's valence–arousal circumplex from raw neural data, and demonstrates causal steering of generated text without modifying model weights.
More >Abliteration-et-Constitutional-Classifier++
White-box activation probing for LLM safety — refusal-direction ablation vs. a
Constitutional-Classifiers++ gate.
Findings — 36-coordinate Cohen's-d signature matches
38,912-coordinate baseline at identical 99.5% OOD accuracy (0.09% of information).
Cross-Model Generalization of Emotion-Vector Activation Monitoring
Causal validation of emotion-vector activation monitoring across 7 models (6 families, 1B–20B); stoic compresses emotion range 63–83% in all models, sham never reproduces the drop but diverges into looping on Gemma 3 1B, confirming specificity to emotional directions with consistent arc recognition across all 7 models.
More >