CK

RésuméCurrent as of September 2026Matches the PDF

Chandhan Saai Katuri

Applied AI Engineer, production LLM systems. San Francisco, CA. hello@chandhan.com, github.com/ChandhanSaai, linkedin.com/in/chandhan-saai-katuri

Download PDF

Experience

Verita AI Software Engineer, San Francisco, CA

  • Designed and shipped a production vision-LLM audit pipeline that analyzes contractor screen recordings against per-project task policies, classifies work hours (AI use, off-task, guideline violations), and produces sealed, priced, recomputable audit ledgers; benchmarked read models across providers via OpenRouter and achieved 80% precision and recall on a golden evaluation set, with 264 sessions sealed in production. Case study
  • Architected the core of an LLM data-annotation platform (Django REST Framework, React): five-role access control with a capability matrix, an audited workflow state machine with calibration routing, screening and qualification gates, an append-only audit log covering all 54 endpoints, and 10+ annotation widget types; expanded the test suite from 210 to 1,239 tests.
  • Hardened the platform's image-ingestion pipeline with SSRF-safe fetching, presigned re-hosting to S3, and asynchronous batched SQS enqueueing that eliminated timeouts on 1,000-image imports; delivered 6,756 processed images to a client with geo-derived metadata backfill.
  • Delivered the second-generation finance-interview rubric verifier for an AI interviewing product in six days, implementing LLM-judge claim validation, shadow-verifier gating, and hard execution time limits; instituted a mutation-testing standard requiring each test to demonstrably fail against a deliberate source defect.
  • Led SOC 2 Type II readiness from 154 to 165 of 173 controls (95%) during an active audit window: executed a pgAudit and storage-encryption migration across 11 production PostgreSQL databases, deployed GuardDuty/CloudTrail/file-integrity monitoring, produced 40+ evidence artifacts, and remediated authenticated penetration-test findings.
  • Developed a multi-scanner application-security pipeline (gitleaks, OpenGrep, Bandit, Trivy, CodeQL) with LLM-assisted triage, reducing 345 raw findings to 33 verified vulnerabilities including nine critical issues (remote code execution, SSRF, IDOR, exposed credentials). Open-source version

Handshake AI AI Data and Evaluation, remote

  • Evaluated frontier image-generation models against calibrated quality rubrics (rendering artifacts, anatomical consistency) to produce preference data for RLHF pipelines; reviewed Kaggle machine-learning codebases for algorithm selection and implementation quality. Case study
  • Selected as a Star Fellow for a program evaluating model reasoning against frontier research publications.

Projects

GPT-2 Transformer Implementation and Training Pipeline PyTorch, CUDA

  • Implemented the complete GPT-2 architecture and training pipeline in PyTorch with bfloat16 mixed precision, torch.compile, fused scaled-dot-product attention, gradient accumulation, cosine learning-rate scheduling with warmup, and BPE tokenization. Project page

Adversarial LLM-Judge Evaluation Framework (Project D3) Python, multi-agent systems, OpenAI API

  • Developed a structured adversarial evaluation framework in which advocate, judge, and jury agents deliberate over competing LLM responses; implemented single-round and iterative multi-round debate protocols with persona-diverse juror voting, per-evaluation token budgeting and cost tracking, and position-swap bias measurement; benchmarked on MT-Bench and AUTO-J, with an AlignBench loader. Case study

Metered API Billing Platform Django, PostgreSQL, React, TypeScript, Docker

  • Engineered a production-grade usage-based billing platform with idempotent event ingestion, tiered invoicing with precise monetary rounding, HMAC-signed webhooks with replay protection, and a database-enforced append-only audit log; shipped customer and operations dashboards as two React/TypeScript applications backed by 145 tests against live PostgreSQL. Case study

Education

University of Maryland, College Park Master of Engineering, Robotics

SRM Institute of Science and Technology B.Tech, Mechatronics (Robotics specialization)

Skills

Languages: Python, TypeScript/JavaScript, SQL, C++

LLM systems: VLM pipelines, RAG, embeddings and vector search (FAISS, Pinecone), LangChain/LangGraph, LLM evaluation (LLM-as-judge, golden datasets, mutation testing), OpenAI/Anthropic/OpenRouter APIs

Backend and web: Django/DRF, FastAPI, React, PostgreSQL, Celery, SQS, REST APIs

Infra and DevOps: AWS (S3, RDS, SQS, CloudWatch, GuardDuty, Elastic Beanstalk, ECR), Terraform, Docker, GitHub Actions, Sentry

ML / DL: PyTorch, Transformers, computer vision (YOLO, ViT), ONNX/TensorRT edge inference

Current as of September 2026. Bullets describe my own work; group and course projects are labelled as such on their pages. The billing test count (145) and the D3 benchmark scope (MT-Bench and AUTO-J) are the verified figures from the repositories.