Software Engineer — AI/ML · Agentic Systems · Full Stack

CS @ UT Dallas ('27, Finance minor). I build autonomous agent systems and the infrastructure that keeps them trustworthy: orchestration, evaluation, and retrieval pipelines that hold up outside a demo.

Currently at Nebula Labs. Open to SWE, backend, and AI engineering roles.

shriramjana.com · LinkedIn · shriram.j.iyengar@gmail.com

Aegis  ·  🔨 currently building
MCP-native security middleware for agent deployments in air-gapped environments. Sits between the model and its tools to enforce policy on what an agent is allowed to call, and writes an audit trail of every invocation.
Python

Trajectory Prediction · 🔨 currently building Predicts where an agent goes next from a few seconds of observed positions. Trained on Argoverse, scored with minADE/minFDE. Emits K candidate futures rather than one averaged path, because an agent at an intersection may turn either way and the mean of those is a curb. Python PyTorch

Sentinel · v1 complete
AI-assisted retrieval and triage for telecom incidents. Combines BM25, dense retrieval, RRF fusion, and cross-encoder reranking to search historical incidents from plain-language outage reports. Evaluated across 2,730 synthetic alarms and 100 frozen operator-style queries, demonstrating that incident relevance and causal-alarm identification require separate objectives.
Python Cohere Qdrant Docker

evalgate
A regression gate for LLM systems. Scores answer quality against a committed baseline and fails CI when a change degrades it — so prompt and model changes get the same scrutiny as code.
Python

Autonomous Research Assistant (ARA)  ·  live demo →
Turns a research question into a cited, multi-source report. A Planner decomposes the question into 3–7 sub-queries, parallel Researcher agents search the live web, and a Synthesizer merges the findings — all streamed to the browser over SSE with per-agent status and a running cost meter. Citations are stable UUIDs, so sources never get mis-numbered when deduplicated. A failed researcher degrades gracefully: the report still completes and notes the gap. Bring-your-own-key, with credentials encrypted per-user via Fernet.
Python FastAPI LangGraph Next.js Supabase

Autonomous AI Prediction Market Trading Bot
Monitors 4,100+ live Polymarket markets, ingesting 120+ news articles per polling cycle to find contracts mispriced against breaking news. Ten LLM analyst personas produce a crowd-consensus probability; Kelly Criterion sizing caps risk at 5% per trade. A circuit breaker halts trading on three consecutive losses or 10% daily drawdown. 130 unit tests across a 5-layer pipeline.
Private repository — architecture walkthrough available on request.
Python FastAPI React

portfolio-rag-chatbot
Retrieval-augmented chatbot answering questions about my background, running in production on shriramjana.com.
TypeScript

Notebook — Nebula Labs  ·  https://dev.notebook.utdnebula.com →
Student-facing AI-powered note platform at UT Dallas. Next.js and PostgreSQL backend with multi-tenant auth and role-based access control.
TypeScript

UTD Trends — Nebula Labs  ·  trends.utdnebula.com →
Course and professor data visualization serving 20,000+ UT Dallas students — aggregates grade distributions and Rate My Professors scores so students can compare sections before registering. Contributed frontend redesigns that lifted engagement 5x, and backend optimizations that cut API latency 30%.
TypeScript Next.js

Python TypeScript Java C++ SQL · React Next.js Node.js FastAPI · LangGraph PyTorch scikit-learn · PostgreSQL Docker AWS GCP