Sanjay Utchula — AI · Backend

§03 Featured work · 2024 → 2026

Six systems. Six different problems.

01 Knowledge graph · Fraud detection

GraphGST

A Neo4j knowledge graph that traces multi-hop GST circular-trading fraud, verified by an Isolation Forest model, and explained by a GraphRAG assistant.

2nd Prize · National Hackathon Team project

cypherMATCH (a:Entity)-[:INVOICE*1..5]->(a)

nodes— entities

detect— circular trade cycles

Ingest

Filings, invoices, and counterparties modelled as a property graph in Neo4j — 1-to-many edges typed INVOICE, GSTIN, DIRECTOR.

Traverse

Cypher pattern-matching over 1–5 hop cycles surfaces suspicious closed loops of invoicing that flat SQL joins miss entirely.

Verify

Scikit-Learn's Isolation Forest scores the subgraph — anomaly signal decoupled from rule-based thresholds.

Explain

A GraphRAG assistant walks the cycle back to an auditor in plain language, citing the exact edges & entities involved.

Neo4jCypherPython FastAPIFlask Scikit-LearnGraphRAG

02 AI governance · On-chain settlement

AgentGuard

A governance layer that puts caps, route restrictions, risk thresholds, and human approval between an AI agent and its wallet — settling verified spend on Algorand.

x402 · USDC · Algorand TestNet Team project

01 · Agent initiates action

02 · Policy caps · routes · risk

03 · Human gate approval threshold

04 · x402 payment USDC transfer

05 · Algorand on-chain record

06 · Verify receipt · audit trail

Why it matters

Autonomous agents ship faster than budget teams can audit. AgentGuard puts the guardrails between the model and the money — inspectable, revocable, and provable on-chain.

My contribution

Policy engine primitives — caps, route filters, risk scoring.
Human-approval routing and audit surfaces.
x402 & USDC settlement integration on Algorand TestNet, verified end-to-end.

Stack

TypeScriptNext.jsHono GroqSupabasePrisma AlgorandPyTealx402

03 Backend systems · AI failure triage

Hookline

A webhook delivery platform with at-least-once guarantees and a two-tier AI failure-triage engine that decides — per failed delivery — whether to retry, throttle, circuit-break, or dead-letter.

96% classification accuracy · +46 points over rules-only Individual project

Triage accuracy · 24-case labeled corpus

Rules only

50%

Rules + Gemini

96%

Verifiable at GET /dev/eval · confusion matrix included

fail 502 · 403 · lying-200

rules deterministic tier

gemini ambiguous only

verdict retry · break · dead

The problem

Webhook failures are ambiguous: an HTML 502 could be a bad gateway or a bug, a 403 is throttling or a dead key, and a 200 body might quietly say "error". Each demands a different response.

Transactional outbox

Business write and webhook enqueue commit in one Postgres transaction — an event can never be lost or orphaned across a crash boundary.

Two-tier triage

A rules engine handles the obvious cases at zero LLM cost. Only the ambiguous residue is escalated to Google Gemini 2.5-flash with structured, closed-enum output. Verdicts cache in Redis.

Backpressure & safety

Redis ZSET delay queue for precise retry scheduling · per-endpoint Resilience4j circuit breakers · Stripe-style HMAC-SHA256 signing so consumers can verify authenticity and reject replays.

Java 21Spring Boot 3.4 PostgreSQL 16Redis 7 Resilience4jGemini 2.5-flash React 18Docker

04 Production case study · AI Internship

Sociovia — SocioEngage

A multi-tenant SaaS for cross-platform social publishing. I contribute to the Instagram automation module and to the wider platform's scheduling, publishing, analytics, and SSO surfaces.

Multi-tenant SaaS Production engineering

sociovia/ig-worker · redis · webhook.ts

// receive comment → route → hand off to DM automation
export async function onCommentWebhook(evt) {
  const tenant = await resolveTenant(evt.recipient_id);
  if (!tenant.enabled) return 204;

  const match = tenant.rules.find(r => r.matches(evt.text));
  if (!match) return 204;

  await igQueue.add('dm.reply', {
    tenantId: tenant.id,
    to:       evt.from.id,
    template: match.dmTemplate,
    gate:     match.followGate,
  }, { removeOnComplete: 1000, attempts: 5 });

  return 204;
}

Instagram automation module

Comment-to-DM, follow-gate, and DM workflow automations built with Meta Graph API and webhooks.

SocioEngage platform

Contributed to cross-platform publishing, scheduling, analytics, and AI-assisted content generation surfaces.

Backend infrastructure

Production-grade scheduling pipelines with Redis queues; multi-workspace data isolation across tenants.

Cross-product SSO & security

Single-use authorization, session isolation, and resolution of critical token vulnerabilities.

Meta Graph APIWebhooks RedisQueues SSOTenant isolation AI content

05 Systems · Concurrency

Java Exam Monitoring System

A JavaFX desktop app simulating live exam-monitoring event processing. O(n) sliding-window algorithm, multithreaded, backed by a ConcurrentLinkedQueue.

Individual project

events · ConcurrentLinkedQueue

window · O(n)

event in window anomaly

Concurrent ingest

Multiple producer threads push proctoring events (focus, click, keypress) into a ConcurrentLinkedQueue without locks.

Sliding window · O(n)

A single consumer maintains a fixed-size window over the event stream — amortized constant per event.

JavaFX surface

Live-updating charts and per-candidate flags, driven from a background thread with correct FX application-thread hand-off.

JavaJavaFX ConcurrencyMultithreading

06 Android · Steganography

PixelCrypt

A native Android app that hides text inside image pixels using Least Significant Bit substitution — bit-level manipulation on android.graphics.Bitmap, exported losslessly as PNG, entirely offline.

Native Android · Java + XML Individual project

plaintext

HELLO

bit stream

01001000 01000101 01001100 01001100 01001111

pixel value LSB flipped carries payload

The technique

Least Significant Bit substitution — the human eye cannot distinguish a colour channel value changing by one, so the last bit of every pixel channel is a free carrier for payload data.

Bit-level on Bitmap

Text → UTF-8 bytes → bit stream, then written into the LSBs of successive android.graphics.Bitmap pixels via direct integer bit-masking.

Lossless export

Bitmap is saved as PNG via MediaStore — PNG's lossless compression preserves every bit, so the extraction round-trip reads the same LSBs back verbatim.

Offline · local · private

All encoding and decoding happens on-device. No network, no cloud, no key exchange. Presented behind a glassmorphism UI built in Java + XML in Android Studio.

JavaXML AndroidAndroid Studio android.graphics.BitmapMediaStore LSB steganography

/ also built

Supporting projects

07 Desktop automation · Python

Nova Voice Assistant

A Python desktop voice assistant with wake-word ("Nova") detection. Speech recognition + text-to-speech, weather & news via OpenWeatherMap & NewsAPI, Wikipedia & YouTube search, system control (volume, folders, shutdown) via nircmd.

PythonSpeechRecognitionpyttsx3 pywhatkitOpenWeatherMapNewsAPI

Sanjay Utchula / engineering knowledge graphs, AI governance, & production backends.