Software, AI & robotics · Graduating May 2027
Mechatronics at IIT Bhilai. I build full-stack apps, AI agents, and ML models, and co-founded Human Slop. I also contribute to open-source robotics tools: 15 merged pull requests across HFlow, Inspect Robots, and Cerulion, plus inspect-robots-wandb, a plugin I wrote and publish on PyPI.
About
Hi, I'm Vamsi. I study Mechatronics at IIT Bhilai, which officially makes me a bit of everything: part mechanical, part electrical, part "can you fix my laptop". I took that personally.
On any given week I'm training a model, shipping a full-stack app, sending a pull request to some robot's software, or arguing with a simulator about physics. Then I close the laptop and go shoot and edit a short film. My browser has too many tabs and I regret none of them.
The full quote goes "jack of all trades, master of none, but oftentimes better than a master of one." I'm betting on the second half. Scroll down for the receipts.
Featured
Open source · Hebbian Robotics, YC S26
HFlow is an open-source SDK for verifying the quality of robot training data. Eight of my pull requests merged between September and October 2026, mostly making robot-log and LeRobot conversion keep data whole or refuse it with a clear reason. The benchmark PR (#367) reported no safe speedup and left the filter graph unchanged.
8 merged PRs · Sep–Oct 2026 · Python, LeRobot, Apache Arrow
Package I maintain · Robocurve, YC S26
Inspect Robots is Robocurve's open-source evaluation framework for physical AI. Its maintainers asked for my Weights & Biases sink to live outside the core, so I published it on PyPI as inspect-robots-wandb, with a 100% coverage gate, CI on Python 3.10–3.13, and trusted publishing. Upstream, 2 of my PRs have merged and 3 are open.
PyPI v0.1.0 · 2 merged PRs · 3 open
Course project · Control theory
Advanced Control Theory course project, simulation only. It reproduces Na et al. (IEEE TSMC 2022) approximation-free active suspension control in MATLAB/Simulink: after fitting two plant parameters to the paper's Case-10 experiment, the unchanged controller lands within about 6% of its error indices (IAE 0.129 vs 0.122). It then tests the claims: the paper's Case-B gains fail a hold-out test, LQR and skyhook baselines match or beat it at equal effort, and an anti-windup extension keeps the wheel on the road over a 50 mm bump.
MATLAB · Simulink · C port timed for a Cortex-M4 · FreeCAD rig · browser lab
Research software · RF · MIT
An open-source, sample-accurate IQ-level emulator for reconfigurable intelligent surfaces: timestamped RIS commands take effect on an exact, logged sample, and controllers run in the loop, deterministically or in real time. Unmodified GNU Radio 3.10 runs through it with bit-for-bit repeatable output. On an AWS c7i host, timed commands with at least 4 ms lead landed on their exact sample in at least 99.8% of cases at 10–100 commands per second. It grew out of the RIS-SIM emulator from my 2025 internship, and was built with Claude Code under my direction.
Python · ~2,100 tests · GNU Radio · ZeroMQ · sole-author preprint draft
Open source
HFlow / Hebbian Robotics (YC S26) / open-source SDK for verifying the quality of robot training data / 8 merged PRs
Most of these PRs make robot-log and LeRobot conversion keep data whole or refuse it with a clear reason.
TRANSFORM_BEHAVIOR_VERSION to 10.camera_frame_stats evidence benchmark: 3 cold 1080p30 checks, 900 decoded frames each, and a 4.897 s median; separated transform timing and provenance, and documented that no repeatable semantics-preserving speedup was demonstrated, so the shipped filter graph stayed unchanged.Link: rel="next", preserved request headers, deduplicated paths, rejected unsafe cross-origin links, and prevented pagination loops.Robocurve (YC S26) builds open-source tools and independent benchmarks for physical AI. Inspect Robots is its MIT-licensed evaluation framework; its CI gates merges on 100% core test coverage, strict mypy, and ruff. / 2 merged PRs / 3 open / author of inspect-robots-wandb on PyPI
max_retries and backoff_s on the LLM agent policy across five wire protocols; HTTP retries honor Retry-After (seconds and HTTP-date) with exponential backoff fallback, and the effective settings are recorded in the eval log.git status failed, making an unknown working tree look clean; now records unknown provenance, with a regression test.eval-set --checkpoint flag.on_eval_error sink hook so sinks holding remote state can clean up when an evaluation aborts; never masks the original exception, notifies each started sink at most once.Open-source Rust robot runtime that records a run and re-executes it against new code (iceoryx2 shared memory, ROS 2 interop) / 5 merged PRs: one CLI fix and four documentation and comment corrections
cerulion ros2 attach --dry-run no longer needs a throwaway workspace. New engine module resolves the root; an exclusively created temp root (create_dir, bounded retries) means a pre-planted directory can never be read as a schema store, with RAII cleanup on every exit path. Hermetic unit tests, engine integration test, and CLI e2e tests with pinned outcome oracles.CERULION_NETWORK kill-switch docs to state the fail-closed arm, verified against source; review surfaced a real gap (topic list never reads the variable), filed by the maintainers as #239.USER_API.md references across 11 files, including runtime error text and its tests.CARGO_MANIFEST_DIR comments in cerulion_cli_engine.Experience
Paid internship / Advisor: Dr. Arzad Alam Kherani / COMET Foundation-sponsored project on Smart Radio Environments
Anti-AI social app for human-written posts
Projects
3-DOF Cartesian gantry that cleans a 90 × 90 cm whiteboard, simulated in ROS 2 Jazzy and Gazebo Harmonic: URDF/Xacro model, ros2_control, a raster path with 10 cm overlap, and coverage analysis from logged data. Built with Gajanand Kumawat.
Kinematic MuJoCo simulation where a prompt such as "sort red, blue, green" is parsed (rule-based for now) into a plan and a pick-and-place sequence, moving from a kinematic gripper to a SCARA arm with analytic inverse kinematics.
Unity simulation of three electropneumatic training kits (sorting, stamping, vacuum transfer): simulated axes, logical PLC I/O, HMI screens, and telemetry record and replay. Not yet connected to a physical PLC.
Razorpay AI Buildathon prototype: chat, voice, and MCP (10 tools) share one bounded commerce core with frozen INR quotes, Razorpay test-mode payments, audit trails, and row-level security.
Wearable exercise recognition (18 exercises plus null) from one-second IMU and video tiles of four unseen subjects, scored by macro-F1. 5th of 232 teams on the final private leaderboard (0.93137; 7th on the public one at close, 0.93825). A fork of goodpjw2008’s public notebook, itself built on notebooks by woominyo and honghanhhh. My additions lifted my reproduction from 0.92497 to 0.93825 public: per-subject whitening, whitened video-continuity link features, ZCA video input, and a multi-run ensemble decode. Built in working sessions with Claude.
Hackathon team project (FortyGuard Hackathon'26, Team HumanSlop): a neighborhood heat-planning scorecard for US planners combining thermal maps, heat-exceedance and streak analysis, vulnerability data, indoor sites, and walking routes.
Team project for Arista's problem at Inter IIT Tech Meet 14.0 (5th place, Dec 2025): client-aware Wi-Fi radio resource management in NS-3, with I/Q-based interference classification (STFT, DBSCAN, Dual-CNN; 16 classes) on a 5-AP, 50-client 802.11k/v topology.
Inverse antenna design: a surrogate network plus KD-tree nearest-neighbour anchoring over about 2,000 simulated patch-antenna S11 spectra, extended to 12 × 12-pixel (144-dimensional) layouts with a separate method-of-moments dataset, plus MEEP scripts for dataset generation.
Crater extraction from lunar DEMs, reproducing Zhou et al. (2018): F1 0.839 on a 16-crater labelled region in Mare Serenitatis (13 true positives, 2 false positives, 3 misses); an optional CuPy path makes preprocessing about 4× faster.
Terrain-only backend API that turns contour maps into pond-site and catchment recommendations through metric projection, elevation-grid interpolation, D8 flow routing, and GeoJSON outputs.
GPU video-effects app in Python, ModernGL, and GLSL: CRT emulation using bundled CRT 3D LUT colour profiles, ASCII edge detection, and 4K export.
In-progress team project (advisor: Dr. Soumajit Pramanik): a two-tower graph neural network (GATv2) that predicts absorption-spectrum peaks from molecular geometry, so far trained and checked on two small molecules (water and ammonia).
Team entries. AIR 200 in 2026. In 2025, team rank 700 (20,000+ registered participants) with a Qwen2-VL 7B and ViT/RoBERTa two-tower model for product price prediction.
Team project: Flutter wellness app with an on-device TFLite mood classifier from the camera, a stress questionnaire, exercises, games, and an expert-contact screen.
Research
Blog
One post per experiment: what I set out to test, the exact numbers I measured, and the parts that did not work.
0.92497 to 0.93825 public, 0.93137 private. What moved the score, a leak that looked like a gain, and why every top team dropped.
Read the write-up ->
Keyword rules blocked 12 of 20 real questions; regex let 10 of 12 jailbreaks through. A typed decision model got both to zero on the same 50 labelled messages.
Read the write-up ->
inspect-robots-wandb v0.1.0 is on PyPI: one W&B run per Inspect Robots evaluation, the full eval spec as config, and every aggregate metric.
Read the write-up ->
Experiment 01: a 1.9M-parameter JEPA world model on an 8 GB laptop GPU — seven minutes of training, emergent physical state, and a planner that scored exactly zero.
Read the write-up ->
Media
A selection from my YouTube channel, from startup stories and engineering work to short-form filmmaking.
Skills
ROS 2 (rclpy, ros2_control, URDF/Xacro), Gazebo Harmonic, MuJoCo, Unity (C#), MATLAB/Simulink, LQR and Kalman filter design, LeRobot datasets
Python, TypeScript/JavaScript, C/C++, Rust (open-source contributions), SQL, GLSL
PyTorch, PyTorch Geometric, OpenCV, NumPy, CuPy, Apache Arrow, Weights & Biases, MEEP (FDTD), CST Studio Suite, NS-3, GNU Radio
FastAPI, Node.js (Fastify), React, React Native (Expo), Flutter, PostgreSQL, Supabase, WebSockets, Docker, AWS (CDK, Lambda, Step Functions, Bedrock), Git, Linux, GitHub Actions, pytest, Ruff, mypy, PyPI trusted publishing
Achievements
Gallery
Resume
One page: open-source contributions, research, projects, and skills.
Contact
Open to robotics software internships and new-grad roles (graduating May 2027), and to research collaborations. I read the code, write the tests, and say plainly what did not work. Email or LinkedIn reaches me.
galabavamsi12@gmail.com / +91 7483047846
Filmmaker / Editor / Director / Writer / IIT Bhilai
I write, direct, shoot, and edit narrative films, music videos, startup content, and social-first stories where culture, technology, and visual mood meet.
Selected Work
Directed by Galaba Vamsi and Harshal Singh. Cinematography by Harshal Singh, Galaba Vamsi, and Aditya Singh. Edited by Galaba Vamsi.
Director, editor, and writer with FPS — IIT Bhilai, developing, shooting, and finishing films under extreme time constraints with a student production team.
Created and contributed to launch and social content for the anti-AI social platform I co-founded, translating a technical product into visual stories for an internet-native audience.
Film Archive
Process
My technical background in robotics, simulation, computer vision, and software gives me an extra edge with cameras, digital tools, visual effects, interactive media, and technical products.
Links
For narrative work, music videos, startup content, social-first reels, and unusual projects that need both visual instinct and technical fluency.
galabavamsi12@gmail.com / +91 7483047846
