AI Engineer · Dallas, TX · open to relocation

Suneel Kalva

I build AI systems that survive contact with production.

Five years of shipping LLM and ML systems end to end: retrieval over 500K documents, fine-tuned open source models, autonomous agents with cryptographic safety gates, and real-time inference for millions of daily transactions.

Get in touchView projects

0+

years shipping production ML

0K+

documents indexed in RAG systems I've built

0%

task success on my incident agent's eval harness

0ms

P95 retrieval latency, tuned down from 800ms

About

The interesting problem isn't training a model. It's trusting one.

I learned to ship AI in one of the hardest environments there is. In banking, a wrong prediction costs real money and every model answers to a regulator. That forced a habit I now carry everywhere: build the evaluation, monitoring, and rollback story before the demo, not after.

The systems I've shipped run at real scale. Retrieval over half a million documents, models trained on ten million records, scoring for two million transactions a day. Lately I've been building agentic systems in the open, including an autonomous incident response agent published on PyPI and a real-time speech translation pipeline that dubs video in the speaker's own cloned voice.

The work I'm proudest of is the part nobody screenshots. Evaluation harnesses, drift monitoring, release pipelines. The machinery that catches problems before customers ever see them.

Open-source projects

Agentic systems I build and maintain in the open.

Mendrift

Autonomous incident response for ML pipelines

source ↗PyPI ↗

An open source agent that handles ML production incidents on its own. It detects drift, runs a bounded diagnosis loop, and proposes a fix. Published on PyPI and listed in the official MCP Registry.

bounded loopdrift detecteddiagnosepropose fixHMAC gatehuman approvesexecute
  • Destructive actions sit behind a single-use, HMAC-signed approval gate in the tool layer. I tested it adversarially against an LLM claiming full authorization with a fabricated token.
  • Incidents survive process death. LangGraph checkpointing pauses before execution and resumes in a new process after human approval.
  • 95% task success on a 19-scenario trajectory harness with a hard zero-ungated-writes assertion.
  • LangGraph
  • MCP
  • Claude
  • MLflow
  • Evidently

Parlecho

Real-time speech translation with voice-cloned dubbing

source ↗

A six-stage pipeline that translates speech and dubs it back in the speaker's own cloned voice, with background music preserved and per-speaker voices intact.

DemucsseparateWhispertranscribepyannotediarizeNLLB-200translateXTTS-v2clone voiceremixtime-align
  • 0.83s median and 1.60s P95 end-to-end dubbing lag on an RTX 4090 under real-time pacing, with queue backpressure included in the measurement.
  • int8 quantization via CTranslate2 cut peak VRAM by 39% and quadrupled GPU throughput with zero BLEU degradation.
  • Cross-lingual voice cloning validated at 0.44 to 0.58 ECAPA cosine similarity against a 0.05 to 0.27 mismatched-speaker baseline.
  • faster-whisper
  • NLLB-200
  • XTTS-v2
  • CTranslate2
  • pyannote

Selected work at U.S. Bank

Fraud investigator's assistant

Analysts at U.S. Bank used to dig through 500,000 policy documents and case files by hand. I built the assistant that reads them instead. Hours of research now take minutes for 200+ investigators, and P95 retrieval latency dropped from 800ms to 180ms.

  • LangChain
  • GPT-4
  • Pinecone
  • AWS

Compliance fine-tuning

Off-the-shelf models don't know AML and KYC. I fine-tuned open source models on the bank's own documentation until they did, raising answer accuracy from 71% to 89%.

  • LLaMA 3
  • Mistral
  • LoRA
  • Hugging Face

Experience

AI/ML Engineer · U.S. Bank

Aug 2024 - Present

The systems that help the bank fight fraud and answer customers, and the infrastructure that makes shipping them safe.

  • Built the fraud team's document assistant covering 500K+ policies and case files. Hours of research reduced to minutes for 200+ analysts.
  • Fine-tuned open source language models on compliance and AML documentation, raising answer accuracy from 71% to 89%.
  • Rebuilt the model release process, taking deployments from three days of manual steps to under an hour.
  • Set up drift monitoring and automated retraining that cut production incidents by 60%.
  • Wrote the LLM evaluation harnesses for hallucination, PII safety, and prompt injection defense, and defined rollout criteria with product and risk stakeholders.

ML Engineer · Fractal Analytics

Aug 2021 - Jul 2023

Consulting for major US banks on fraud, churn, and customer service models.

  • Predicted churn across ten million customers for a Fortune 100 client. The campaigns built on it cut attrition by 18%.
  • Cut false fraud alerts by 25% on two million daily transactions with no drop in catch rate.
  • Built the NLP pipeline that routes five million customer complaints to the right team.
  • Moved model serving onto Kubernetes for three times the throughput at 30% less cost.

ML Engineer · Cognizant

Jun 2020 - Aug 2021

Where I started: credit risk models and the data pipelines that feed them.

  • Replaced a rule-based credit scorecard with statistical models that were 12% more accurate.
  • Served real-time scoring for 500K+ daily transactions in under 200 milliseconds.
  • Built daily ETL pipelines processing a million-plus customer records.

Skills

LLM systems & agents

  • LangGraph
  • LangChain
  • RAG
  • MCP servers
  • tool calling
  • fine-tuning (LoRA/PEFT)
  • guardrails & evals

Machine learning

  • Python
  • PyTorch
  • TensorFlow
  • XGBoost
  • Scikit-learn
  • SQL

Inference & speech

  • CTranslate2
  • int8 quantization
  • streaming inference
  • Whisper
  • XTTS voice cloning
  • speaker diarization

MLOps & infrastructure

  • MLflow
  • SageMaker
  • Docker
  • Kubernetes
  • Airflow
  • Spark
  • Pinecone / FAISS
  • AWS / Azure / GCP

Education

M.S. in Computer Science · University of North Texas · Denton, TX

Aug 2023 - May 2025

Let's build what's next.

I'm open to AI engineering roles where the hard part is making it work in production, at early-stage teams and large product companies alike. If you're shipping serious AI, or just want to compare notes, my inbox is open.

suneel.kalva190700@gmail.comDownload résumé