Software Engineer · AI Developer
I build AI-native software — LLM agents, retrieval & evaluation pipelines, and the backend systems that make them safe to ship.
Software Engineer Intern @ Settlyfe · M.S. Applied Data Science @ San José State University ('27) · San Jose, CA
- 🤖 Agentic systems — LangGraph ReAct agents with typed tools, streamed tool calls, and multi-agent fan-out → Agentic Software Platform, OTTO, Le Bistro
- 🔎 Retrieval + evaluation — embeddings, hybrid vector search, guardrails, and eval harnesses that gate releases → FilmFinder, LedgerLens, vision-eval-harness
- ⚙️ Backend & data systems — FastAPI / Go / Spring Boot services on Postgres & Redis — containerized, CI-gated, observable → Baseline, Digital Intelligence, BookTable
The pattern across all of them: deterministic core, LLM at the edges, and a test, eval, or guardrail in front of every model output.
|
An in-browser IDE where an AI agent is a first-class collaborator, not autocomplete. Monaco editor + WebContainer sandbox in the browser; a LangGraph ReAct agent with six Zod-typed tools (list / read / write / delete / search / install) edits the project live over an SSE event stream; Yjs CRDT multiplayer editing with live cursors and share-link roles; GitHub OAuth; every agent run and tool call persisted.
|
The accounts-payable inbox that reads invoices — and never auto-pays. PDF → SHA-256 dedup → parse / OCR → LLM classification + field extraction (16-node RocketRide pipeline) → a deterministic Postgres brain (three-layer dedup, guard rules, anomaly stats) → human review queue. The same graph runs on a laptop (Ollama) or in the cloud (Gemini); a 43-doc synthetic corpus and eval results are committed — 100% clean / 94.3% scanned field accuracy, 0 payable-queue leaks.
|
|
Ctrl+F for game film — search a soccer match in plain English. A vision LLM captions frames once, offline; at query time FastEmbed + Qdrant hybrid search fuses a dense vector with action and time filters, so "saves in the last 10 minutes" is a sub-second query. Two Lyzr agents sit behind flags with plain-Python fallbacks, a circuit breaker, and a code-enforced grounding gate; the eval harness scores 12/12 at rank 1 (MRR 1.00) — and caught a real agent bug on day one.
|
CI for model quality — a statistical release gate for ML models. Runs vision and LLM eval suites, then blocks the merge only when a candidate is genuinely worse than a checksummed golden baseline: paired exact McNemar, Benjamini–Hochberg across suites, PSI / KS drift, bootstrap CIs. Async LLM runner with token-bucket rate limiting and a SQLite response cache; parallel, quarantining transcript ETL; Typer CLI with an exit-code contract; GitHub Actions gate.
|
|
A biometric platform that learns your normal and flags when your body drifts from it. Encrypted-at-rest ingest (FastAPI + Fernet), Hampel + zero-phase Butterworth filtering, personal median / MAD baselines, and four detectors — z-score, CUSUM, IsolationForest, and a per-user PyTorch autoencoder — feeding an offline-first React PWA whose IndexedDB outbox syncs itself when connectivity returns. Full stack in one
|
Restaurant table reservations, end to end. Spring Boot REST API secured with Spring Security (JWT + Google OAuth2) over MongoDB — customers, restaurant managers, and admins each get their own surface; React + Material UI frontend wired through an Axios service layer; Docker Compose for the full stack and GitHub Actions CI/CD deploying to AWS EC2 / S3.
|
How the Agentic Software Platform fits together
flowchart TB
subgraph B["Browser"]
ED["Monaco editor · file tree"]
YJ["Yjs CRDT<br/>y-monaco · y-webrtc"]
WC["WebContainer sandbox<br/>npm install · dev server · preview"]
CH["Agent chat panel<br/>streamed tokens + tool cards"]
end
subgraph S["Next.js 16 · API routes"]
AUTH["NextAuth v5 · GitHub OAuth<br/>getProjectAccess → role"]
RUN["/api/agent/run<br/>SSE stream"]
AG["LangGraph ReAct agent<br/>list · read · write · delete · search · install"]
end
DB[("Prisma · SQLite<br/>projects · files · runs · messages · shares")]
PEERS(("collaborators"))
ED <--> YJ
YJ -. "P2P" .- PEERS
ED --> WC
ED -- "debounced save" --> AUTH
AUTH --> DB
CH --> RUN
RUN --> AG
AG --> DB
AG -. "file_change events" .-> CH
| Project | What it is | Stack |
|---|---|---|
| OTTO | Voice-first web-agent fleet: a planner splits one spoken ask into up to three read-only subtasks, fans them out to parallel browser agents, and a synthesizer speaks the answer — one typed event stream drives the whole ops console; mock mode replays it with zero backend (hackathon build) | Python · FastAPI · asyncio · React |
| Le Bistro | Mobile ordering app where an LLM manages the cart through validated, structured tool calls; provider strategy for Claude / OpenAI / Groq, output sanitizer with fuzzy item resolution, Vitest suite | Expo / React Native · Hono · Drizzle + SQLite · Zod |
| Digital Intelligence | Go telemetry collector → Redis Streams → PySpark feature jobs; bearer auth, per-tenant rate limiting, OpenTelemetry tracing + Prometheus metrics, Helm charts, k6 load tests | Go · Redis · PySpark · Helm |
| DocV Behavioral Prototype | FastAPI scoring service with Pydantic session contracts and a reproducible synthetic dataset generator for bot-vs-human document-verification sessions | Python · FastAPI · Pydantic · pytest |
Contributor to rocketride-org/rocketride-server — an open-source AI pipeline engine with a C++ core and Python-extensible nodes:
- #2001 · merged — first unit tests for
rocketlib'sdepends.py; removed the duplicate cache tests - #1998 · merged — declare
pyjwtso the published engine image starts - #2002 · merged — refresh the MCP tool catalog instead of caching it once
| Area | Tools I reach for |
|---|---|
| Languages | Python · TypeScript / JavaScript · Go · Java · SQL |
| AI / LLM | LangChain · LangGraph · OpenAI / Gemini / Claude / Groq APIs · Ollama · FastEmbed & sentence embeddings · Qdrant · pgvector · Chroma · RocketRide pipelines · evals, guardrails & grounding gates |
| Backend | FastAPI · Next.js (App Router / API routes) · Spring Boot · Hono · REST · SSE / WebSockets · asyncio |
| Data & storage | PostgreSQL / Supabase · Redis Streams · MongoDB · SQLite · Prisma · Drizzle · SQLAlchemy / Alembic · PySpark |
| ML / analytics | PyTorch · scikit-learn · SciPy · pandas · statistical testing (McNemar, Benjamini–Hochberg, PSI / KS) |
| Frontend | React 19 · Next.js 16 · Tailwind · Expo / React Native · Monaco · Yjs CRDTs · PWA / Service Workers |
| Infra & DevOps | Docker / Compose · GitHub Actions · AWS (EC2, S3) · Vercel · GitHub Pages · Helm / Kubernetes · OpenTelemetry · Prometheus |
- 🏢 Software Engineer Intern @ Settlyfe — Java / Spring Boot microservices, PostgreSQL, Docker on AWS
- 🎓 M.S. Applied Data Science @ San José State University · graduating May 2027
- 🔭 Recently shipped: LedgerLens's cloud deploy + weekly digest pipeline, the
vision-eval-harnessCI gate and Pages reports, and my first merged PRs to RocketRide - 🧭 Exploring next: multi-agent topologies (planner / coder / reviewer), codebase RAG on pgvector, and tracing + cost budgets for agent runs
📫 jenilsavalia058@gmail.com · LinkedIn · San Jose, CA
Open to Software Engineering and AI Engineering roles.


