Software Engineer (AI) Intern · Voice Games
AI Engineer · Full Stack · Multi-Agent Systems
Tushar Agrawal

Tushar Agrawal

Software Engineer (AI) Intern at Voice Games, building autonomous multi-agent AI systems and full-stack systems. Currently focused on developing intelligent workflows, retrieval pipelines, and developer tools that solve practical real-world problems.

01 / Experience

A track record of building.

May 2026 — Present
Remote
Voice Games

Software Engineer (AI) Intern

Engineered TraceLens, an AI-assisted frontend intelligence platform for browser audits, trace analysis, and bottleneck detection.Contributing to UI systems, debugging workflows, and gameplay performance optimization for live production games on Jest.com.

PlaywrightLighthouseTypeScriptNode.jsCLIPerf Eng
Aug 2025 — Present
Jaipur, India
LNMIIT · Technology Council

General Secretary

Leading 7+ science & technology clubs and overseeing PLINTH, LNMIIT's 3-day national tech fest. Heading a team of 100+ members driving strategy, execution and cross-team collaboration for 40+ events and 1,000+ participants.

LeadershipEvent OpsCross-teamPublic Speaking
July 2023 — June 2027
Jaipur, India
The LNM Institute of Information Technology

B.Tech, Computer Science

Coursework across OOP, DSA, DAA, DBMS, OS, Computer Networks and Computer Organization.

DSAOSDBMSNetworksOOPPython
02 / Projects

Things built with intent.

echo · live call · vapi
connected · 00:42p95 · 1.85s
recruiter · "tell me about your RAG work"
echo · "Built Echo — a dual-channel agent grounded on a 33-pair golden set, 98.2%…"
/ 01

Echo

Autonomous AI persona that represents you to recruiters.

A dual-channel AI agent, live voice telephony & web chat ,that represents a developer's full professional profile to recruiters in real time. Embeds resume and GitHub data as 3072-dim vectors into Pinecone, orchestrates GPT-4o Mini across a Vapi voice pipeline (Deepgram Nova-3 STT → ElevenLabs Turbo TTS), and closes the loop with autonomous Cal.com interview booking, all within a sub-1.3s conversational response cycle.

  • RAG pipeline with 3072-dim OpenAI embeddings indexed in Pinecone, achieving 98.2% groundedness on a 33-pair golden Q&A set scored by a GPT-4o judge.
  • Voice pipeline via Vapi · Deepgram Nova-3 STT at 99.1% accuracy · GPT-4o Mini at ~390ms latency · P95 response 1.85s
  • End-to-end Cal.com booking with 6/6 confirmed test bookings, which is hardened against timezone offset, slot-starvation and webhook routing failures.
Next.js 15VapiPineconeDeepgramElevenLabsGPT-4o MiniCal.comTypeScript
// graph.intentsync
1,284 nodesconfidence · 86%p50 · 28ms
/ 02

IntentSync

GraphRAG-powered repository intelligence engine for engineering teams.

A local-first CLI tool that indexes a repository's commit history, PRs, issues, and file relationships into a multi-store GraphRAG knowledge base, then answers natural language queries against it. TypeScript monorepo with 6 isolated packages, backed by a BullMQ async worker pipeline.

  • Triple-store retrieval — ChromaDB (vector), PostgreSQL (structured), Neo4j (co-change graph) fan-out for richer grounding than single-store RAG.
  • File co-change coupling via Neo4j surfaces non-obvious architectural dependencies that static analysis tools miss entirely.
  • Answer Confidence Scoring combines Gemini 2.5 Flash's self-assessed groundedness with computed vector similarity scores to output a verified confidence percentage (e.g. 86% HIGH) per response, making retrieval quality auditable rather than opaque
TypeScriptNeo4jChromaDBPostgreSQLRedisBullMQGemini 2.5Docker
tracelens · live
Audits
lcpok
tbtok
clsok
hydrationok
pipeline · run #4821
42 runs/min
/ 03

Tracelens

AI-assisted frontend performance intelligence.

Built at Voice Games to automate browser audits, trace analysis and bottleneck detection for the core engineering team. Powers live latency optimisation on the Jest.com platform.

  • Automated analysis pipeline integrating Playwright, Lighthouse, hydration & bundle intel.
  • CLI-driven workflow for multi-run audits, regression compare and consolidated reports.
  • Standardised the team's pre-release performance evaluation harness.
TypeScriptPlaywrightLighthouseNode.jsCLI
codonova · planner.agent.py
01 async def plan(spec):
02 graph = build_dependency_graph(spec)
03 for task in graph.topo():
04 code = await coder.generate(task)
05 tests = critic.run(code)
06 if not tests.passed:
07 code = repair_loop(code, tests) # 76→96%
08 return ship(graph)
● tests · 96% passretries · 2/5agents · 4 active
/ 04

Codonova

Autonomous AI development framework.

A multi-agent AI framework that autonomously plans, generates, tests, debugs and self-corrects production-ready software codebases orchestrated over a graph-driven pipeline with persistent contextual memory.

  • Graph-driven orchestration with Neo4j + ChromaDB for dependency-aware task execution.
  • Automated TDD self-correction loop (Pytest/Jest) that lifted generation success from 76% → 96%.
  • Fault-tolerant LLM layer: key rotation, exponential backoff, provider fallback.
FastAPIReactNeo4jChromaDBDockerLLMs
03 / Skills

The stack I reach for.

/ 01

Languages

CC++JavaScriptPython
/ 02

Frameworks & Libraries

React.jsNext.jsNode.jsExpress.jsFastAPI
/ 03

Databases & Tools

MongoDBNeo4jChromaDBDockerGitGitHub
/ 04

AI & Agents

Agentic WorkflowsContext EngineeringRAGStructured OutputsMulti-agent
/ 05

Backend & APIs

REST APIsJWTAuthCloudinaryWebhooks
/ 06

CS Foundations

DSAOOPDBMSOSComputer NetworksDAA
04 / Contact

Let's build intelligent
systems together.

Working on something ambitious in AI, devtools or infrastructure? Drop a message or book a slot directly on my calendar.

0/2000
Or skip the back-and-forth

Book a meeting
directly.

Pick a slot that works for you. Cal.com handles the timezone math.

© 2026 · Tushar Agrawal