Backend · Distributed Systems · AI Security

Pavan Kumar Gannoju

Backend & Platform Engineer

I build backend and platform systems focused on correctness, security and recovery. My projects cover transactional ledgers, replicated storage, Kubernetes admission control and bounded AI agents, with reproducible tests and documented failure behavior.

About Me

I like the parts of backend engineering that fail quietly and are hard to debug.

Background

My work sits where backend correctness, platform security and AI safety overlap: distributed consistency, failure recovery, admission control and LLM guardrails. In these systems a mistake is not a typo. It is a production incident.

I use automated tests, CI, fault injection, load experiments and fuzzing to investigate how systems fail. Each project documents the checks performed, the results and the limits of that evidence.

The three areas are the same problem at different layers: a system should behave predictably when its inputs, its network or its model cannot be trusted.

Engineering Strengths

  • Transactional backends with idempotent writes and replayable state
  • Kubernetes admission control and policy enforcement
  • Replicated systems tested with crash and fault injection
  • LLM guardrails evaluated with labelled and adversarial fixtures
  • CI pipelines that run fault injection, fuzzing and load tests

Technical Skills

Everything listed here appears in a project or internship on this page.

Languages

JavaGoPythonRustSQL

Backend

Spring BootFastAPIREST APIsKafkaJWTgRPC

Cloud & Infrastructure

DockerKubernetesAzureGitHub ActionsTerraformHelmKustomize

AI Systems

LLM GuardrailsAgent SafetyPolicy GatesModel Artifact SecurityEvaluation Harnesses

Security & Validation

cosignTrivyFuzzingProperty-Based Testingk6PorcupineFault Injection

Datastores

PostgreSQLMySQLSQLiteRedisFAISSQdrant

Experience

Where I worked, what I actually did, and the documents to back it up.

March 2025 – November 2025 | Remote | Part-Time

AI Intern → Research Team Lead

Prodigal AI Technologies Pvt. Ltd.

Built & Delivered
  • Built the AI workflow pipeline and backend orchestration layer that the team's research prototypes ran on, covering prompt routing through output validation.
  • Promoted from intern to team lead. Led a team of four, set the technical direction and ran the architecture reviews on LLM safety.
  • Defined the research direction on zero-shot voice cloning with Mixture-of-Experts and on adversarial tamper-resistance, which produced two research papers.
May 2025 – September 2025 | Remote (Jersey City, NJ) | Part-Time

Gen AI/LLM Intern

Digital Nexus AI

Built & Delivered
  • Built backend REST APIs in Python and FastAPI for the core GenAI product features.
  • Connected retrieval, generation, validation and service-layer logic into a modular pipeline that the team could maintain and extend.
  • Shipped the integration with tests and wrote handover documentation for the engineers who took it over.

Featured Projects

Each project links to its tests, validation results and known limits.

01

LedgerLine

Reliable transactional ledger for idempotent transfers and replay-safe financial projections.

Verified candidate 43c8a0c · CI checked 1 Nov 2026. Feature branch; not merged into main.


Java 21, Spring Boot, PostgreSQL, Kafka, Spring Security (OAuth2 resource server), Docker, k6, Azure

323–330
transfers/s · Oct 26 Azure runs
341–362 ms
p95 · same Azure runs
0
HTTP failures · same runs

  • Implemented balanced integer-minor-unit postings, principal-scoped idempotency and a transactional outbox feeding replayable Kafka projections.
  • Added durable poison-event quarantine, restricted reconciliation views and a separate read-only incident worker; hosted CI checks the Java services, worker and both service images.
  • Earlier Azure validation recovered projections after consumer and Kafka outages in 20.5 and 9.3 seconds, and restored 253,652 ledger postings with all 11 table fingerprints matching.

Evidence & limitations

The metrics describe two 150-second Azure runs on 26 Oct 2026 with Premium SSD v2 storage (3,000 IOPS), not the current candidate. Code and storage changed; no code-only gain is claimed. The later local hot-account workload missed its original latency limits despite a passing many-account run. Current candidate CI passes; a 24-hour soak and production operation remain unverified. OAuth2 validates access tokens; it does not provide login or token issuance.

02

ModelGate

Kubernetes admission control for signed images, model-artifact checks and namespace-scoped workload policy.

Verified candidate e9af94f · CI checked 1 Nov 2026. Feature branch; not merged into main.


Go, Kubernetes admission webhooks, CRDs, cosign, kind, Helm, Kustomize, Azure ACR

cosign
digest-bound image signatures
kind
admission and lifecycle checks
CI
trust, fuzz and worker checks

  • Enforced namespace-scoped policy with digest-bound signature verification, workload restrictions and fail-closed behavior on trust failures.
  • Added serving readiness, bounded graceful termination and a durable proposal-only worker outside the admission decision path.
  • Hosted CI passes real cosign integration, envtest, kind, Helm, fuzz and worker checks, including Linux FIFO rejection regressions.

Evidence & limitations

Policy exemptions and installation exclusions are documented explicitly. The local observation stopped after 15 hours; the original 24-hour gate is incomplete. A distinct-image load control regressed. Azure ACR evidence records image publication, not a managed deployment. Production operation and host-level high availability remain unverified.

03

SentinAL

Safe Windows desktop-agent prototype with validation gates, audit logs, and post-action checks.


Python, FastAPI, SQLite, Windows automation, local/LLM-assisted intent handling

96.7%
task success (95% CI 91.7–98.7%)
95.26%
held-out accuracy
66/66
fuzz cases blocked

  • Supported actions are validated before execution and, where a check exists, verified against OS state afterwards.
  • Reached 96.7% task success (95% CI 91.7–98.7%), 95.26% held-out accuracy and 83.68% out-of-distribution accuracy, with 88.45% of requests resolved on the fast path.
  • Backed by 1,333 tests passed, 2 skipped, at 86.75% coverage (October 6, 2026 CI run), plus a 66-test fuzz suite with a 100% block rate.
  • Led a four-person team on this B.Tech thesis project.

Evidence & limitations

Windows-only prototype with bounded capabilities. Figures come from the project's own evaluation sets and test suites, and the 100% block rate is over the 66-test fuzz suite.

04

QuorumKV

Replicated key-value store with durable writes, Raft replication, and failure-behavior testing.

Verified candidate 0d13601 · CI checked 1 Nov 2026. Feature branch; not merged into main.


Go, gRPC, HashiCorp Raft, WAL, mTLS, fault injection, Porcupine checks

100/100
WAL crash trials passed
mTLS
peer and client authorization
CI
Linux / Windows worker checks

  • Built the WAL, store adapter and observed-execution harness around HashiCorp Raft; histories retain acknowledgements, retries and uncertain write outcomes.
  • Added authenticated peer/client transport, bounded requests and exact-state snapshot/restore checks, with real process failover and certificate-rotation tests.
  • Separated durable diagnostics from consensus and repaired checkpoint publication with crash/concurrency regressions. Hosted core, fault, race, worker and checkpoint jobs pass.

Evidence & limitations

The 100 WAL trials exercise populated process-kill recovery, not hardware power loss. Fault checks cover finite observed histories, not proof of every execution. The earlier endurance run failed after about eight hours; checkpoint repair and smoke checks do not establish a completed 24-hour run. Default diagnostics use deterministic offline analysis; live model quality remains unverified.

05

Gatekeeper

LLM guardrail gateway for prompt-risk detection, policy decisions, and safe request routing.


Python, FastAPI, Redis, LLM security, classifier evaluation

Up to eight
configured detectors

  • Up to eight detectors contribute to each assessment, depending on configured assets. More than 1,800 automated tests.

Evidence & limitations

No current independent accuracy or latency headline is published; see the repository's evaluation notes.

06

Agentic-OS

Rust operating-system prototype for capability-scoped interfaces, policy, and auditability.


Rust, QEMU, kernel development, serial tests, systems safety

Rust
native kernel work
QEMU
validated runtime

  • Kernel written in Rust with typed capability interfaces, and no raw pointers without a reason.
  • QEMU boots, serial output validates and host tests confirm, reproducible from a clean checkout.

Evidence & limitations

Work in progress. Runs in QEMU.

Additional Projects

Smaller projects in applied ML, data and audio.

Enterprise RAG System

RAG pipeline over policy documents for an FP&A backend. Swappable vector backends (NumPy, FAISS, Qdrant), schema-validated output, streaming, rate limiting and property-based tests.

  • Replaced serial per-chunk embedding with bounded concurrency: a 181-chunk ingest dropped from 378.7 s to 48.1 s (7.87×).
  • Load-tested with concurrent HTTP traffic and found the ceiling is model generation, not the API layer. Fixed a mid-stream connection-reset bug that only appeared at scale.
  • 343 Python tests at 98% branch coverage, with CI green on both Python versions.
PythonFastAPIFAISSQdrantOllamaDockerNext.js

DDS — Driving Decision System

Driving simulator with decision logic: OBD-II telemetry, XGBoost + SHAP behaviour analytics, Frenet-frame path planning and IDM car-following, streamed live to a 3D browser HMI.

PythonFastAPIXGBoostSHAPThree.js

SoundIntelligence

Captures system audio on Windows, classifies it with FFT + YAMNet, and writes a live Equalizer APO profile, so the EQ adapts as the audio changes without manual presets.

PythonWASAPIYAMNetQt

Research & Leadership

Research direction and team leadership at Prodigal AI.

AI Security Research

At Prodigal AI I defined the research direction on two problems the team ran into, and wrote a paper on each. Both papers are currently unpublished.

  • Zero-shot voice cloning with Mixture-of-Experts architectures
  • Adversarial defenses for systems whose model output cannot be trusted
Research experience letter

Engineering Leadership

Promoted from AI Intern to Research Team Lead at Prodigal AI (March – November 2025, part-time). Led a team of four, keeping research moving toward prototypes that could ship.

  • Set the direction for the AI workflow architecture and the team's approach to LLM safety
  • Ran architecture reviews that changed the design before it was built
  • Built the workflow pipeline and orchestration layer the team's prototypes ran on
Recommendation letter

Certifications & Achievements

Security, AI and engineering credentials. Cards with a certificate open in a viewer.

AI-ML Engineer, NSQF Level 5

NSDC / PMKVY

National skills qualification in AI and machine learning

SAP Code Unnati

SAP

Machine learning, IoT, computer vision, ABAP and SAP HANA

Smart India Hackathon 2023

MoE, Govt of India

Grand Finale participant

Coding Club President

Leadership

2024 – 2026

More credentials

Click any to open the certificate

GitHub

Source code, tests and validation notes for the featured projects.

@pavann19

Each featured project has its own repository with source code, tests and validation notes, so the numbers on this page can be checked against the code.

Python TypeScript Java Go Rust React & Next.js
View GitHub Profile

Contact

Open to backend and platform engineering roles in India and Germany. Send a message and I'll get back to you.