AI/ML Engineer · B.Tech Mathematics and Computing, IISc Bangalore (graduating 2027)
I build AI systems that have to prove they work: models checked against their own math,
LLM evaluation that fails the build on a regression, and agents that finish only when the tests pass.
Open to 2027 full-time roles and internships: AI/ML Engineer · ML Engineer · Applied AI · AI Automation
| Result | Where |
|---|---|
| 10.7x lower latency and 42.6x less GPU memory than a Transformer | Ericsson internship, Mamba wireless model on AWS A10G |
| Groundedness collapse (0.57 → 0.00) blocked in CI on every push | AIRE, LLM/RAG evaluation engine |
| 8/8 adversarial tool calls rejected, 0 failed tool calls | repo-engineer, verified coding agent |
| 98.3% parity accuracy where baselines stay at chance (≤56.3%) | s4-to-mamba, Mamba-3 reproduced in PyTorch |
| Top 100 national finalist | Amazon ML Challenge 2026 |
LLM evaluation RAG CI quality gate · Python, pytest, GitHub Actions, Gemini API
Trace-based evaluation for LLM and RAG systems: run a system once, store validated traces, then score them, compare versions and attribute every regression to the cases that caused it. The aire-gate CLI turns a comparison into a CI check.
The repo's own CI plants a "fluent but uncited" version and the gate blocks it on every push: groundedness 0.57 → 0.00 and ungrounded failures +22, even though retrieval recall rose 0.73 → 0.91. A live Gemini 2.5 Flash run raised abstention 0 → 1.0 at 15 s per case, and the report prices that trade-off instead of hiding it.
Code · Live demo · Report (PDF) · Case study
agents tool calling sandboxing · Python, pytest, Gemini API
A coding agent where the LLM only proposes JSON tool calls. A deterministic state machine (plan → act → observe → verify) validates every call against a schema, path-jails it to the workspace and applies AUTO/ASK/DENY permissions. A task is done only when the tests pass.
4/4 bug-fix, test-writing and refactoring tasks with 0 failed tool calls, 2/2 with a live Gemini planner on the same runtime, and 8/8 path-escape, command-injection and malformed calls rejected.
Code · Live demo · Report (PDF) · Case study
deep learning state space models GPU benchmarking · PyTorch
Minimal S4D, Mamba, Mamba-2 (chunked SSD) and Mamba-3 in pure PyTorch, with 22 tests checking that the parallel and recurrent forms agree, and training benchmarks from 1K to 32K tokens on a 6 GB laptop GPU.
Reproduced the Mamba-3 state-tracking result: complex transitions solve parity at 98.3% while real-transition SSMs and a Transformer stay at chance (≤56.3%). On selective copying, LTI S4D collapses to 6% while selective SSMs reach 67%.
Code · Live demo · Report (PDF) · Case study
AI/ML Research Intern · Ericsson India · Bengaluru · May 2026 - Jul 2026
Benchmarked WiMamba (a Mamba wireless model) against a Transformer on AWS A10G GPUs: 10.7x lower latency (30 ms vs 327 ms) and 42.6x less GPU memory. Evaluated transfer to an unseen geography (+16.6% composite score) and parameter-efficient fine-tuning that kept 98.5% of full fine-tuning performance with 75% of parameters frozen.
Machine Learning Engineer · Alignerr and micro1 · Remote · May 2025 - Aug 2025
Wrote reasoning traces and tool-use data for LLM post-training (Alignerr), and built model evaluation, validation and data-preprocessing workflows in Python with TensorFlow and scikit-learn (micro1).
- Amazon ML Challenge 2026: Top 100 national finalist; led Team AutoML.
- Oracle Cloud Infrastructure Certified Professional: Generative AI and Data Science
LLM systems
RAG LLM evaluation and regression testing tracing tool calling agent runtimes Gemini / OpenAI APIs
Sci-Tech Coordinator for Rhapsody Fest at IISc: ran three inter-college gaming tournaments, a coding contest with 200+ teams and a 500+ member Discord community. Off the clock it's Jujutsu Kaisen and Dragon Ball Z. The Gojo in me trusts the process; the engineer in me checks the logs anyway.



