AI Engineer

Sreekar Reddy Portfolio

Applied AI Engineer at Pretian Squared

Applied AI Engineer (ex-IBM, building internal AI platforms at Pretian Squared) whose portfolio features nine projects, including ISM CyberRAG, a team-built university capstone RAG system with published RAGAS faithfulness and answer-relevancy scores, and Cortex, a code-review tool built as a council of six specialist agents.

RAG PipelinesAgentic WorkflowsLLM EvaluationMulti-Agent Systems

Location

North Sydney, NSW, Australia

Companies

Pretian SquaredCity QuokkaAI Campeden.pmIBM

Credentials

  • Master of Artificial Intelligence, University of Technology Sydney (2025–2026, in progress)
  • B.Tech in Computer Science Engineering (IoT), Jain (Deemed-to-be University) (2017–2021)

Skills

RAG PipelinesAgentic WorkflowsLLM EvaluationMulti-Agent Systems

Notable

  • UTS Dean’s List 2026
  • IBM Think40 Champion Learner — Gold (2023, 2024)

What makes it work

A breakdown of the choices that make this portfolio stand out.

1

RAG quality reported as named, specific metrics

Citing faithfulness (0.68 to 0.84) and answer-relevancy (0.72 to 0.91) scores, measured on the same 100-question set across versions, for the CyberRAG project gives a concrete, evaluable measure of retrieval quality instead of a vague "built a chatbot" claim.

2

Multiple small, distinctly named projects instead of one flagship

Nine featured projects, each with its own name and write-up — a RAG system, a verification layer for AI-written code, a recruitment research agent, an MCP prompt library, a browser-native AI dev environment, a multi-agent review council, a P2P video app, a local-first 3D knowledge graph and the site itself — show range across the stack rather than one narrow demo.

3

A specialist-agent structure named explicitly

Describing the code-review tool as a "council" of six specialized agents, rather than one generic assistant, shows deliberate multi-agent architecture thinking, not just a single wrapped LLM call.

What AI Engineers can take from this

Specific, actionable tips to apply to your own portfolio — no generic advice.

  • Report actual evaluation metrics (faithfulness, relevancy scores) for any RAG system you've built — it's far more convincing than describing it as "accurate."

  • Ship several small, distinctly named agent projects rather than one large unfinished one — range across the stack demonstrates versatility.

  • If a project uses multiple cooperating agents, name their distinct roles explicitly — it shows you're designing an architecture, not prompting a single model.

Projects

Sreekar Reddy's featured projects read like release notes: dated, broken down to one-line architectures, and the flagship RAG system's repo keeps a results table with the one sprint target it missed, by 0.012. For an engineer whose pitch is reliability, the miss is part of the evidence.

ML ModeladvancedOne university term, three sprints (UTS AI Studio, Autumn 2026)

ISM CyberRAG: Evaluated RAG over the Australian Information Security Manual

PythonFastAPISupabase pgvectorBM25Cross-encoder rerankingGroq Llama 3.1 8Bnomic-embed-textRAGASClearMLDockerGitHub ActionsHugging Face Spaces

A cited, natural-language search system over the 25 PDFs of the December 2025 Australian Information Security Manual. The baseline used fixed 1,000-character chunks and vector-only search. It was replaced by 643 chunks cut at ISM control boundaries, each carrying control IDs and applicability metadata. BM25 and pgvector results are merged through reciprocal rank fusion and rescored by an ms-marco-MiniLM-L-6-v2 cross-encoder, then Llama 3.1 8B answers with control-ID citations. A two-stage guardrail (a regex pre-filter, then a calibrated rerank-score threshold) refuses out-of-scope questions before the model is called. Built as a UTS capstone by a team of three, with roles rotating each sprint.

The same fixed benchmark ran across every version, so each gain can be traced to the change that produced it. Security guidance is a domain where a confident wrong answer does damage, so refusal behavior was engineered and calibrated (a ClearML sweep set the threshold) rather than added afterwards.

  • RAGAS on a fixed 100-question set: faithfulness 0.68 to 0.84, answer relevancy 0.72 to 0.91

  • Control-aware chunking: roughly 900 fixed-size chunks replaced by 643 ISM-aware ones

  • Server-Sent Events Pipeline Explorer streams all seven stages live, from pre-filter to generation

  • Results table keeps the one missed target (answer similarity, short by 0.012)

TooladvancedHackathon build (OpenAI Codex Hackathon Sydney), extended after

Scout: Verification Layer for AI-Written Code

Next.js 16React 19TypeScriptOpenAI APIMCP TypeScript SDKGitHub APIZodTailwind CSSVercel

An MCP-native reviewer aimed at the failure modes of coding agents: fake imports, README/code drift, weak tests and patches that look right but do not apply. Specialist lanes (Hallucination, Spec Drift, Test Theater) produce findings. A judge layer dedupes them and labels each one confirmed, likely or speculative. A patch tournament then scores conservative, idiomatic and robust repair candidates. An execution gate applies each diff in a temporary workspace, with secrets stripped from the environment, and disqualifies any patch that fails to apply.

Scout doesn't take model output on trust. It separates model judgment from deterministic proof and makes a patch apply before it can win, which is the discipline teams need once coding agents write a real share of the code.

  • Seeded seven-bug benchmark with measured recall and precision, not 'review vibes'

  • Execution gate: candidate diffs must apply in a temp workspace before a patch can win

  • Official MCP TypeScript SDK server with five tools, plus resources and prompts

  • Context-budget telemetry: inspected files, estimated tokens, cache keys

Web AppintermediateNot stated (dated December 2025)

Cortex: Multi-Agent Code Review Council

Next.js 16TypeScriptOllama Cloud APIMonaco EditorZustandTailwind CSS v4

Code pasted into a Monaco editor is reviewed by a council of six AI agents. Each has its own role and system prompt: Architect, Sentinel, Optimizer and others. A Moderator merges the findings into a ranked verdict, and a Verifier checks each claim against the actual code. The agents run on Ollama Cloud models from 120B to 671B parameters. Security review goes to the Sentinel, which runs DeepSeek v3.1 (671B).

Cortex shows orchestration design rather than a single wrapped prompt. Retries use exponential backoff and each agent has its own failover, so if one agent hits a rate limit the review still returns partial results instead of failing.

  • Six role-specific agents plus a Moderator and a Verifier that challenges claims

  • Model per role: DeepSeek v3.1 (671B) handles the security lane

  • Per-agent failover and exponential backoff instead of all-or-nothing reviews

  • Access code gate on the live demo to prevent API abuse

Open SourceintermediateNot stated (dated January 2026)

MCP Prompt Library: npm Package, CLI and MCP Server

TypeScriptNode.jsMCP SDKnpmZodGitHub Actions

115+ prompts, workflows and coding standards for AI-assisted development, organized into 8 categories and distributed three ways. The npm package @esreekarreddy/ai-prompts exposes getPrompt, searchPrompts, listPrompts and composePrompts. There is an ai-lib CLI, and a TypeScript MCP server with 15 tools for Claude Desktop, OpenCode and Cursor. Five guided workflow chains cover new features, bug fixes, refactoring, security hardening and production launch.

It's a working tool shipped the way real libraries ship, with a typed API, a CLI, an MCP interface and CI. A reviewer can install it and try it in a minute.

  • One prompt corpus, three interfaces: npm API, CLI, MCP server

  • 15 MCP tools, including intent-based prompt suggestions

  • 147 automated checks in GitHub Actions

Ready to build your portfolio?

Follow the AI Engineer roadmap — skills, projects, and timeline to get hired.

AI Engineer Roadmap