Control System Engineer at Toshiba. Electrical engineer by degree, wrote a lot of software, came back to hardware, then fell into interpretability. Robotics is where it all meets.
Everything I've done keeps circling one problem: if a model hasn't seen something before, how will it act? That's all. It's a very simple problem.
You can't teach a model every single situation. A chatbot can say "I don't know". A robot can't — it has to do something. And the models that end up in robots won't be calling a cloud and waiting for an answer; they'll be small, sitting on the hardware, with very little compute to check themselves. That's the part I want to work on: embedded software, embedded hardware, and models that behave safely and predictably on stuff they were never trained on.
Next: master's in robotics / ECE, Fall 2027. I don't want to do the same old thing again and again. I want to understand systems and build new ones.
- An Assurance Is Not Evidence: Answerability Claims Change What Agents Assert — tells an LLM agent "this is answerable" and measures how it starts asserting unsupported answers instead of abstaining. Being submitted to an AAAI 2027 workshop.
- Decomposed Swarm Fusion (DSF): A Hypothesis for Multi-Axis Conceptual Fidelity Through Functionally Specialized Lightweight Networks — single-author hypothesis paper. Instead of one big model holding every axis of meaning, a swarm of small specialised networks sharing one latent space. Predictions + falsification criteria, no experiments yet. Zenodo preprint, DOI 10.5281/zenodo.20360400.
- Capability degradation mapping — experiment. Remove MLP neurons from a small language model task by task: when does a capability fall off a cliff, and what breaks along with it? Pilot + v2 done, write-up paused.
- esp32-signal-fault-mlp — experiment. ~2 KB int8 neural net that tells a healthy signal from a faulty one on an ESP32. Simulated in Wokwi, synthetic data, and it says so.
- SLM- — experiment. Built a small language model from scratch, just to see what everyone was talking about.
Work projects
- DR Collector — IEC 61850 — needed a cheaper in-house disturbance recorder instead of buying one. Built it in C on libIEC61850: relay trips, stores the record, the workstation pulls the COMTRADE files over MMS automatically, with reconnect handling.
- Wire tracer → cable schedules — a model that traces wires through substation drawings and produces cable schedules, running on local hardware. A second model scores its confidence, and when it's unsure it asks an engineer instead of inventing an answer. That rule had to be built in; rechecking everything yourself is doing the whole problem again. (Internal, can't share.)
My own
- CONNECTOR-02 — my invention: a one-piece wire-clamp connector for relay panels, designed in Fusion 360. Indian provisional patent filed May 2026, in my name.
- Conductor — routes my work between Claude Code, Codex and Gemini CLIs depending on the task.
- GeoOSINT-GLOBE — side project: 3D globe with seven live feeds — satellites, aircraft, ships, earthquakes, wildfires, weather, cameras.
- PDF-Studio — PDF editor in Flask, with tests.
- event-intel — for capturing contacts and notes at conferences.
- Open Classroom — multi-agent AI classroom from a PDF or a topic.
Learning the internals
- ARENA — working through the interpretability curriculum. Own ResNet-34 matches torchvision's predictions; transformers next.
- Advance Auto Parts (intern, 2024) — first time working with models for real. They couldn't get enough test data. Expected a one-week job; it took six. CTGAN had no idea which values could repeat and which couldn't, so we built a rule layer on top of it: 10,000 rows in, 100,000 coherent ones out.
- Accenture (Oct 2024 – Mar 2025) — put in financial analysis. Not my thing. Left.
- APC Edutech (co-founder, Mar – May 2025) — education startup teaching Python and ML. It didn't work out, so I moved on.
- Toshiba (May 2025 – now) — joined in Transformers, came second in an internal ideation competition on digital substation monitoring, moved to control & relay panels / substation automation. SAS architecture, IEC 61850, a lot of talking to customers and vendors, and the guy who tries the new AI tools.
Education: B.Tech EEE, VNR VJIET Hyderabad, 2024 (8.25/10, First Class with Distinction) · IIT Madras BS Data Science, Foundation level
Tools: Python · PyTorch · C · IEC 61850 (MMS/GOOSE) · Modbus · MATLAB/Simulink · Fusion 360 · TypeScript · PostgreSQL