ML systems and agentic AI. CS @ UC San Diego (2028).

I build systems that survive contact with production: inference pipelines, LLM agent infrastructure, and the eval/observability layer that tells you when they break. Currently splitting time between GTM engineering at a GPU cloud, applied medical imaging ML, and whatever hackathon is happening that weekend.

Open to freelance work. Details at the bottom.

LinkedIn · Portfolio · lakshgoyal06@gmail.com

• GTM Engineering @ GMI Cloud Outbound and revenue infrastructure for a GPU cloud: enrichment pipelines, automated workflows, and the data plumbing behind them.
• Applied ML @ a medical imaging startup Segmentation, depth estimation, 3D reconstruction, and structured clinical output from a single photo.
• ML Systems @ Engineers for Exploration (Qualcomm Institute) Multi-terabyte bioacoustic pipelines and domain-shift robustness for wildlife monitoring.

Compresses prompts along two orthogonal axes so the savings multiply: token-space (LLMLingua-2 + LongLLMLingua + a from-scratch AttentionRAG implementation, since the paper ships no code) and model-space (TurboQuant as a drop-in DynamicCache subclass, ~4-bit KV). Stacked with LCLM latent-context compression for ~60x total KV reduction while beating the fp16 F1 baseline on LongBench. Voice-first demo over Deepgram Nova-3, served on warm Modal A100s.

PyTorch Transformers Modal FastAPI Next.js

4K LOC. Scores accuracy, per-run cost, and latency percentiles, then ranks agents on production trade-offs instead of raw accuracy (cost-adjusted accuracy, efficiency score). Async-safe per-node usage attribution via contextvars that survives asyncio.gather. Ray Tune + ASHA for prompt/model HPO. Shipped as a package: Typer CLI, FastAPI leaderboard, GitHub Actions CI with mypy --strict, OIDC PyPI publishing.

Python LangGraph LiteLLM Ray Tune FastAPI

Geospatial simulation, not a notebook. A 7-stage preprocessing pipeline turns farm boundary polygons into 11 ELMFIRE-ready rasters on a canonical 30m EPSG:5070 grid, pulling live LANDFIRE and USGS 3DEP data behind SHA-256-keyed caching with exponential backoff. Runs 8-bearing fire-spread ensembles, then scores 24 candidate firebreak layouts on a 6-weight composite of protected-area safety, risk reduction, and per-segment cost. 92 tests, mypy --strict, 100% type coverage.

Python rasterio GeoPandas SciPy Pydantic ELMFIRE

Led 11 engineers through a 10-week Agile SDLC (5 sprints, pair programming, weekly role rotation) to ship a Scrum/Kanban/XP tool on Cloudflare Pages with D1, session auth, and a 5-stage CI pipeline gating every PR.

Vanilla JS Cloudflare D1 Vitest Playwright GitHub Actions

More in the pinned repos below.

Four hackathon wins across ML infra, defense tech, and scientific computing. Highlights:

• 1st place, Loop Engineering Hackathon. Vigil, an AI on-call engineer with zero standing credentials (Best Use of Akash)
• 1st place, Vibe Coding Active Matter Hackathon. Agent-based T cell swarming simulation with chemotaxis and receptor desensitization
• 2nd place, BowCapital Defense Hackathon. Pylon, a counter-UAS sensor mesh, now incubated at Bow Capital
• Published through the Lumiere Research Program, top 10% of ~200 submissions
• 2nd place, World Robotics Olympiad India. Autonomous pipeline inspection robot

I take on a limited number of contracts. Best fit:

• LLM and agent systems: RAG pipelines, tool-calling agents, cost and latency optimization, eval harnesses
• ML pipelines: data ingestion, training infra, model deployment
• Full-stack builds where the hard part is the ML, not the CRUD
• GTM engineering: enrichment pipelines, outbound automation, and the data plumbing behind them

Reach me at lakshgoyal06@gmail.com.