ceenaAbout
A bit about me
I'm an early engineer at GRiDD Technologies, where I built GNet Connect from the ground up — a B2B transportation platform spanning reservations, documents, partner discovery, support workflows, and internal operations.
I focus on AI-assisted operational systems that convert unstructured business inputs into structured, reviewable workflows — across the full stack, from React/Next.js UIs to backend APIs, SQL/Redis data systems, and retrieval pipelines.
After hours I go under the hood: training sparse autoencoders on a ~1B GPT I built from scratch, steering models with their own features, running Jacobian-lens interpretability research on a 3B model, and benchmarking TensorRT-LLM inference. B.S. in Computer Science, UC Riverside (2023). The case studies below walk through each system in depth.
Experience
Where I've worked
Full-Stack Engineer · GRiDD Technologies / GNet Connect
2024 — PresentEarly engineer who built GNet Connect from the ground up — a B2B transportation platform for reservations, documents, partner discovery, support workflows, and internal operations.
- Built and shipped full-stack product features across React/Next.js, TypeScript, backend APIs, SQL/stored procedures, Redis/RediSearch, and internal admin tools.
- Designed AI-assisted operational systems built around human review, uncertainty surfacing, and feedback loops — including an AI email intake pipeline for reservations and a Gmail support agent.
- Worked directly with the CTO, CEO, support team, and customers to turn operational problems into data models, UI flows, backend behavior, and shipped product.
Case Studies
What I've built
Full write-ups for each project: the problem, the design decisions, and what I owned.
- J-Space — Cultivating Internal Reasoning
Ongoing research: can fine-tuning change what a 3B model computes internally — its higher-order reasoning space — not just what it says? A Jacobian-lens instrument, a pre-registered falsification bar, and an Atomic workflow built to refuse its own conclusions until the internal evidence exists.
- Steering a GPT with Its Own Features
Trained sparse autoencoders on a ~1B GPT I built from scratch, found that meaning lives in superposition, and injected facts at inference with no weight changes — turning Michael Jackson into a physician. With a live steering demo.
- Inference Bench
A benchmarking API for LLM inference serving: toggle TensorRT-LLM configs (quantization, batching, speculative decoding, KV-cache, chunked prefill), run load tests across chat, coding, and embeddings profiles, and compare throughput and latency — with a live dashboard.
- AI Email Intake for Reservation Workflows
Turning unstructured customer emails into structured reservation drafts, with intent classification, schema validation, missing-field review, post-review learning, and account-level memory.
- Gmail Support Agent
Treating a support reply as a context-assembly problem: classify the issue, gather account and reservation context via MCP tools, retrieve resolved precedents, and ground a draft for human review.
Skills
Tools I work with
Frontend
Backend
Data
AI / Retrieval
Tools / Platforms
Contact
Let's talk
I'm open to interesting problems and good conversations. The fastest way to reach me is email.