TarunGudapati

AI-assisted deployment risk review · MVP

Aegis Release Guard

A working MVP that turns a git diff into cited release risks, focused test suggestions and a rollback checklist.

Result — LLM analysis with a deterministic fallback: runs end to end with no external AI required, xUnit + Vitest suites, CI and Docker in the box.

Role
Design & Development
Stack
.NET 10 · React · TypeScript · xUnit
Year
2026
  1. Git diff
  2. Deterministic pre-analysis
  3. LLM structured analysis
  4. Evidence validation
  5. Risk findings
  6. Recommended tests
  7. Rollback plan

Problem — releases fail in the gap between "the diff compiles" and "the diff is safe." Reviewers skim large changes and miss destructive migrations, loosened auth or secrets slipping into a commit.

What I built — a working public MVP that turns a unified git diff into a transparent risk score, findings tied to exact files and lines, focused test suggestions and a recovery checklist.

Architecture — analysis runs behind an IAnalysisProvider interface. The default provider is deterministic and local: it flags destructive database operations, weakened security controls, credential-like literals, configuration changes, broad release scope and production changes without visible tests. An OpenAI-compatible provider implements the same interface for LLM-structured analysis; the deterministic engine remains the fallback, and model claims are checked against evidence from the diff.

Tech — a .NET 10 Minimal API, a responsive React and TypeScript interface, xUnit and Vitest coverage, plus CI and Docker assets. It is decision support, not a claim that automated analysis can approve a release.

GitHub repository ↗← Back to the full portfolio

Tarun builds systems that ship.

I build production AI systems with .NET, Azure and Angular — LLM features that cite their sources, fail safely, and ship through pipelines that can be rolled back.

View AI projects
RésuméGitHubLinkedInContact

Sweep the strings · turn sound on whenever you like

Software Engineer — AI Systems, .NET & Azure

Hyderabad, India · Remote / worldwide

AI engineering is more than a model call. Evidence, evaluation, fallbacks and deployment are part of the product — because a model is only useful when people can rely on it.

Currently building

  • AI in production

    Shipping LLM features inside a live .NET finance platform — structured outputs, prompt caching, and queries written for the database that actually runs them.

  • Public AI work

    Release-risk review, grounded retrieval and evaluation gates — public repositories with tests and stated limits.

  • Azure delivery

    CI/CD, observability and rollback paths that carry AI changes to users without surprises.

01 / Selected work

Selected work

A flagship AI build, public experiments, production software, and a small collection of sites I make for people.

Featured AI Engineering

The flagship: a working release-risk review MVP with public code, tests, CI and Docker.

  1. Aegis Release Guard — MVP

    AI-assisted deployment risk reviewMVP

    Aegis Release Guard

    A working MVP that turns a git diff into cited release risks, focused test suggestions and a rollback checklist.

    LLM analysis with a deterministic fallback: runs end to end with no external AI required, xUnit + Vitest suites, CI and Docker in the box.

    • .NET 10
    • React
    • TypeScript
    • xUnit
    Read the case studyView repository

AI Engineering Experiments

Public lab work. Each repository states exactly what works today and what is deliberately left out.

  1. TraceRAG — Concept

    Citation-first retrieval scaffoldConcept

    TraceRAG

    A self-checking .NET retrieval baseline that answers from local evidence, cites exact chunks and refuses unsupported questions.

    Every answer is backed by document- and chunk-level citations; unsupported questions return "insufficient evidence" rather than a guess.

    • .NET 10
    • Minimal API
    • Retrieval
    • xUnit
    Read the case studyView repository
  2. IncidentSight — Concept

    Multimodal incident-triage scaffoldConcept

    IncidentSight

    A safety-first FastAPI baseline that extracts log evidence, inspects image metadata and keeps every remediation step human-approved.

    36 tests over a typed, validated pipeline; every remediation step is gated behind human approval and nothing runs automatically.

    • Python
    • FastAPI
    • Pydantic
    • Pytest
    Read the case studyView repository
  3. ModelLedger — Concept

    Offline LLM regression gatesConcept

    ModelLedger

    A working TypeScript CLI scaffold that compares recorded model outputs and fails CI when quality, latency or cost gates regress.

    Deterministic quality, latency and cost gates that fail CI with a non-zero exit code the moment a recorded output regresses.

    • TypeScript
    • Node.js
    • Zod
    • Vitest
    Read the case studyView repository

Selected Work

Production engineering.

  1. Safe Financials — Shipped

    AI inside production finance softwareShipped

    Safe Financials

    Claude API features shipped inside a live .NET / SQL Server financial platform — prompt caching, structured outputs and queries written for the production database.

    LLM features live in production — Claude API with prompt caching and structured outputs, EF Core queries written for the live SQL compatibility level, and fallbacks when the model fails.

    • .NET
    • EF Core
    • SQL Server
    • Claude API
    • Angular
    Read the case study
  2. AmpleLogic — Shipped

    Enterprise software engineeringShipped

    AmpleLogic

    Backend, frontend and deployment work on enterprise software for regulated-industry workflows.

    • .NET
    • Angular
    • SQL Server
    • CI/CD
    Read the case study

Portfolios

Sites I design and build for photographers — a frame as considered as the pictures. Both live.

  1. Live homepage of sandeepsanka.com, a fashion photography portfolio

    Fashion photographer · HyderabadShipped

    Sandeep Sanka

    A cinematic, image-first site for an established Hyderabad fashion photographer whose editorial and campaign work includes Allu Arjun, Vijay Deverakonda, Rashmika Mandanna, and houses such as RWDY.

    Live at sandeepsanka.com — a full-bleed, motion-paced editorial site built with Next.js, TypeScript and GSAP.

    • Next.js
    • TypeScript
    • GSAP
    Read the case studyVisit live site
  2. Live homepage of sujana.zip — “Frames that feel like cinema” over a black-and-white portrait

    Celebrity & portrait photographer · BLR–HYDShipped

    Sujana

    A cinematic, film-still portfolio for a freelance photographer shooting celebrity portraits, actor portfolios, brand campaigns, and pre-weddings between Bengaluru and Hyderabad.

    Live at sujana.zip — a monochrome, film-still portfolio built with Next.js and TypeScript.

    • Next.js
    • TypeScript
    • CSS
    Read the case studyVisit live site

02 / About

About

Portrait, drawn from live data — the loom remembers your movement.

A software engineer specializing in practical AI integration and the systems around it.

My day-to-day work crosses .NET services, EF Core and SQL Server data layers, Angular and TypeScript interfaces, Azure delivery, and LLM-powered product features. I care about grounded outputs, explicit failure modes and what happens after the demo reaches production.

My production experience includes fintech software at Safe Financials and enterprise systems at AmpleLogic. The public projects below show the same approach in the open: working code, tests and documentation.

I also design and build personal sites. The Portfolios collection is that side of the practice: a live photographer’s site and an editorial personal portfolio, composed as custom work rather than a template.

Based
Hyderabad, India
AI
LLM integration · RAG · evaluation · guardrails
Engineering
.NET · C# · Angular · TypeScript · SQL Server
Delivery
Azure · CI/CD · observability · rollback design

03 / Capabilities

Capabilities

The parts of AI development that turn a promising prototype into software a team can inspect, test and operate.

04 / Contact

Build AI people can trust.

Open to full-time roles and product engineering work — especially teams shipping AI features that have to hold up in production.

Full-time roles · Product engineering · Hyderabad · Remote / worldwide · tarungudapati30@gmail.com

Let's talkGitHubLinkedIn