Parth Auti

Software engineer, CS new grad (December 2026). AI agent reliability and applied ML, benchmark-backed open source. At T-Infosystem I built the Bedrock summarization layer with a validation guard behind an automated maintenance-reporting platform covering 400+ client sites. At UNC Charlotte's CharmLab I'm co-authoring a manuscript on LLM-judge failure modes, targeting a NeurIPS 2026 workshop. Open source: Reflight (a flight recorder for AI agents), ChainCheck (claim-level hallucination detection at 94% precision, shipping on PyPI), CostDNA (GraphSAGE for untagged AWS spend), Bourse (Rust matching engine, zero-alloc hot path). Open to SWE / ML infrastructure / systems roles.

• University of North Carolina at Charlotte — B.S. in Computer Science

  Aug 2025 – Expected Dec 2026

  Charlotte, NC · GPA 3.7 / 4.0 · Chancellor's List (Spring 2026) · Dean's List (Fall 2025)

  Coursework: Machine Learning · Deep Learning · Data Mining · Operating Systems · Computer Networks · Database Systems · Cloud Computing · Data Structures & Algorithms

• Manipal Academy of Higher Education (MAHE) — Computer Science (transferred to UNC Charlotte)

  Aug 2023 – May 2025

  Manipal, India · GPA 3.63 / 4.0

• Charlotte Machine Learning Lab (CharmLab) — Undergraduate ML Researcher

  May 2026 – Present

  UNC Charlotte · advised by Prof. Minwoo Lee

  • —Drafting a manuscript on LLM-judge failure modes. Judge misses on HaluEval-QA are high-confidence, invisible to NLI/judge disagreement (AUC 0.48), and concentrate on hallucinations built from the reference passage itself (49% missed vs 31%, p<0.01).
  • —Built a fully reproducible analysis pipeline — all 99 manuscript statistics verified against committed artifacts. Targeting a NeurIPS 2026 workshop submission.

• T-Infosystem — Software Engineering Intern

  Mar 2026 – Jun 2026

  Remote

  • —Built uptime, response-time, SSL-expiry, and incident-detection services for an automated maintenance-reporting platform covering 400+ client sites. AWS Lambda, SQS, EventBridge, DynamoDB, SES.
  • —Developed the Amazon Bedrock summarization layer that generates monthly client reports. Added a validation guard that rejects unsupported figures and falls back to deterministic templates when validation fails.
  • —Created a 40-case evaluation suite for summary accuracy and factual grounding. Pass rate rose from 64% to 87% through prompt and validation improvements.

• UNC Charlotte — AV Technician

  May 2026 – Present

  Charlotte, NC

  • —Live event production: audio and video for campus events up to 800+ attendees — setup, live mixing, on-the-fly troubleshooting.

• Starbucks — Barista, then Barista Trainer

  Oct 2025 – Apr 2026

  Charlotte, NC

  • —Promoted to Trainer ahead of standard tenure. Onboarded new hires on beverage standards, POS, and food safety.
  • —Ran the bar through 200+-transaction morning rushes while carrying a full CS course load.

• Reflight — A flight recorder for AI agents.

  2026

  Records every model call, tool call, token, and dollar an agent spends. Three lines of code to instrument. Replays are byte-for-byte identical to the live run, verified with the network hard-blocked, and finish in about 7 ms at $0.00. On a scheduling agent, a 15-line replay assertion caught 15 out of 15 failures that passed every tool-level check. An LLM judge caught 1 to 5 of them, depending on how the prompt was worded. One command turns a failed run into a pytest regression test.

  Python · pytest · OpenAI · LangGraph · OpenTelemetry · Docker · github.com/pauti04/reflight · live

• ChainCheck — Claim-level hallucination detection for LLMs.

  2026

  Flags the exact unsupported sentence instead of scoring the whole response. 94% precision at 64% recall on HaluEval-QA (n=500, gpt-4o-mini judge, balanced 50/50), chosen after benchmarking four detection methods head-to-head. A 60 ms local NLI pre-filter cuts latency 19× on the clear-cut cases and only escalates ambiguous claims to the judge. Ships on PyPI as a drop-in OpenAI proxy that tags risky responses in headers. A GitHub Action wraps the same core and blocks merges when a PR description doesn't match the diff.

  Python · FastAPI · PyTorch · HuggingFace · OpenAI/Anthropic · github.com/pauti04/chaincheck · live

• CostDNA — Behavioral attribution for untagged cloud spend.

  2026

  On a typical AWS account, 40 to 60 percent of spend is untagged and invisible to FinOps tools. CostDNA infers ownership per resource with a GraphSAGE GNN trained on CloudTrail behavior, IAM patterns, and VPC topology, then writes the tags back. Existing tools go from explaining half the bill to about 95%. End to end from raw cost report to per-team breakdown in 90 seconds, no data leaves the AWS account. I also caught label leakage in two published benchmarks (Azure 2.6M VMs, Philly 117K jobs) by auditing my own results — the honest number is published next to the inflated first-cut one.

  Python · PyTorch (GraphSAGE) · Next.js · Terraform · AWS · GPT-4o · github.com/pauti04/CostDNA · live

• Bourse — Order matching engine in Rust.

  2026

  Price-time-priority matching engine over a length-prefixed binary TCP protocol. Zero allocations on the hot path, proven by a custom allocator harness. Cache-padded lock-free MPSC gateway feeds a single-writer matcher, with Acquire/Release ordering and Miri validation in CI. Byte-exact WAL replay plus snapshot recovery; group commit is 187–245× faster than fsync-per-record. Current benches on M-series release: in-process round-trip ~225 ns, TCP RTT p50 ~78 µs / p99 ~307 µs, matcher walks 1000 levels in ~94 µs.

  Rust · tokio · Miri · lock-free MPSC/SPSC · CRC32C WAL · github.com/pauti04/bourse · live

• NetPulse — BGP anomaly and route-leak detector.

  2026

  Public reproducible benchmark on RIPE RIS archive data: 7 out of 7 labeled historical incidents detected, 0 false negatives, runs in about a second, no API keys. The live monitor taps the RIPE RIS global feed (~1,800 updates/sec); in a 70-second sample it ingested 149,246 announcements. RPKI validation runs at ~43 µs per call against 859k VRPs — a longest-prefix-match trie replaced a linear scan and moved a 43 ms operation to 43 µs. Hosted HTTP API deploys to any Docker host.

  Python · FastAPI · DuckDB · RIPE RIS · fly.io · github.com/pauti04/netpulse · live

• Dispatch — An AI-curated morning brief for developers.

  2026

  A daily brief in the shape of a newspaper — ink on cream, gold accents, double-ruled masthead, issue numbering. Role-first onboarding streams the first headline over Server-Sent Events in about 800 ms. Solo-built end to end: React + Vite frontend, Express + Postgres (Neon) server, gpt-4o-mini generation, Resend for delivery, Sentry and PostHog for observability, Expo for the mobile shell. 41 passing tests, Lighthouse Best Practices 100, zero accessibility violations in the audit.

  React · Vite · Express · Postgres (Neon) · gpt-4o-mini · Resend · Expo · github.com/pauti04/dispatch

• RasoiBot — A pantry-aware recipe assistant for Indian cooking.

  2025

  The smallest thing on this page. I cook a lot of Indian food and kept rewriting the same paneer recipe from memory. RasoiBot is that recipe, but with serving-size scaling and a grocery list. Pantry-aware set-intersection over a hand-curated index — the streaming feel is a UX trick, not a model.

  JavaScript · Next.js · Vercel · github.com/pauti04/RasoiBot-clean · live

languages

Python · Rust · TypeScript · Go · C++ · SQL · Java · Bash

ml / ai

PyTorch · HuggingFace · Transformers · embeddings · LLM agents & evals · RAG · LangGraph · MCP · GraphSAGE / GNNs · pandas · NumPy · NLI · hallucination detection

systems

Tokio · async Rust · lock-free data structures · WAL + snapshot recovery · FastAPI · DuckDB · Postgres · Redis · BGP / RIPE RIS · RPKI

cloud & infra

AWS (Lambda, SQS, EventBridge, DynamoDB, Bedrock, SES) · Azure · Docker · Kubernetes · Terraform · GitHub Actions · Fly.io · Vercel · Render

frontend

Next.js 15 · React 19 · Tailwind v4 · server components

languages

English (professional) · Hindi (native) · Marathi (conversational)

• AWSSolutions Architect – Associate · SAA-C03
• AWSDeveloper – Associate · DVA-C02
• AWSCloud Practitioner · CLF-C02
• AzureDevOps Engineer Expert · AZ-400
• AzureAdministrator Associate · AZ-104
• AzureAI Engineer Associate · AI-102
• AzureFabric Analytics Engineer Associate · DP-600
• Oracle CloudOCI Foundations Associate

• 2026Chancellor's List — UNC Charlotte — Spring 2026
• 2025Dean's List — UNC Charlotte — Fall 2025
• 2024Smart India Hackathon 2024 — 3rd Place (MAHE) — IoT precision agriculture: ESP32 sensor nodes, Firebase, ML irrigation recommendations, GSM/SMS alerts for low-connectivity rural areas
• 2023JEE Advanced 2023 — qualified — Top ~1.3% of 1.4M+ candidates (India's national engineering entrance exam)

• —Full-time SWE, ML infrastructure, or systems roles — graduating Dec 2026
• —Charlotte, NC — remote or relocation, US or international
• —AI agent reliability · LLM tooling and evals · low-latency systems · GNNs · cloud/platform engineering
• —Interviewing now

parth.auti@gmail.com · github.com/pauti04