AI Engineer @ TitanCloud · Open to full-time · mid-2026

Rojan Dahal

I build the systems that gate, validate, and govern other AI systems.

AI Engineer @ TitanCloud · MS Data Science @ Gannon · GPA 4.0
5+ years shipping AI in production. Fintech, manufacturing, agentic systems.

[ Scroll the pipeline ↓ ] [ Read resume.pdf ]
01 / Rules

L1 — what passes the first filter

Reliability matters as much as accuracy.

Five years ago I was building fraud-detection systems for banks in Nepal — signature verification, liveness detection, document classification. Systems where a wrong prediction has real consequences for real people. That wired something into how I approach ML.

Since then I've pushed in two directions. Deeper on research: a peer-reviewed paper on using CNNs to automate powder-bed leveling in metal 3D printing, presented at NAMRC 53 / MSEC 2025. Broader on engineering: a RAG chatbot on GCP Cloud Run with local LLMs, vector DB, Redis, FastAPI; a multi-tenant SaaS with conversational AI that cut support inquiries by 70%; and the Gatekeeper at TitanCloud that gates a Bedrock VLM extraction agent.

The through-line: I care about what happens after the model is trained. Deployment, latency, uptime, real usage. Anyone can run a notebook. I want the thing to actually work in production.

NOW
AI Engineer at TitanCloud. Building production AI on AWS Bedrock + Step Functions.
NEXT
Open to full-time ML Engineer / AI Engineer roles from mid-2026. US-based, remote-friendly. F-1 OPT.
OPEN
Agentic systems, MCP, AI safety, MLOps, document AI, computer vision.
  • 5+ years
  • 1 peer-reviewed paper
  • 3 countries
  • MS Data Science
  • GPA 4.0
02 / Signals

L2 — what I actually reach for

The tools that survive contact with production.

Agentic AI & LLMs

What I reach for when the problem is language, tool-use, or reasoning.

  • Amazon Bedrock (Claude Sonnet 4 / Haiku 4.5)
  • Claude API
  • Model Context Protocol (MCP)
  • LangChain
  • vLLM
  • RAG
  • Bedrock Guardrails
  • Structured-JSON outputs
  • Prompt engineering
  • LLM evaluation

ML & Computer Vision

Five years of training, optimizing, and shipping models — most of them vision.

  • PyTorch
  • TensorFlow
  • ONNX Runtime
  • Transformers
  • CNNs
  • EfficientNet
  • YOLOv5
  • OpenCV
  • DocAligner
  • CLAHE
  • OCR
  • Fine-tuning
  • Liveness detection

Cloud & MLOps

Where the model actually has to behave — across three clouds.

  • AWS — Lambda · Step Functions · SageMaker · S3 · EventBridge · OpenSearch Serverless · RDS
  • GCP Cloud Run
  • Azure — IoT Hub · Cognitive Services · Functions
  • Docker
  • Kubernetes
  • GitHub Actions
  • Azure DevOps
  • DataDog
  • Cloudflare (Tunnels, Pages)

Languages & Data

Day-to-day tools. Python for almost everything; TypeScript for surfaces like this one.

  • Python
  • TypeScript
  • Java
  • SQL / PostgreSQL
  • Redis
  • MongoDB
  • Vector DBs

Production engineering

The boring, important parts most ML resumes skip.

  • FastAPI
  • REST APIs
  • JWT / auth proxies
  • Multi-tenant SaaS
  • CI/CD
  • Canary deploys
  • Sub-100 ms latency systems
  • Rate limiting
03 / Inference

L3 — the work that the rest of the site is downstream of

Four roles. One through-line: production AI that has to behave.

01

Titan Cloud Software

AI Engineer · Mar 2026 — Present

The Gatekeeper — a cost-control filter in front of a 3-agent Bedrock IDP pipeline.

  • Owned the Gatekeeper end-to-end — a 4-layer hybrid filter (L0 DocAligner ONNX + CLAHE → L1 rule pre-filter → L2 keyword scorer → L3 EfficientNet-B0 CNN) that sits in front of the Bedrock VLM. Reframed it from a doc-type router to a cost-control gate, because the dominant cost wasn't misrouting — it was garbage documents reaching the VLM.
  • Designed a 3-agent serverless IDP pipeline on Step Functions — Gatekeeper → Primary Extraction → Validation — where each agent reaches data only through MCP servers (RAG, DB, knowledge graph). The MCP boundary keeps the agent–data contract clean and lets us swap the KG from a PG adjacency list to Neptune without touching prompts.
  • Built Intelligent Prompt Routing (IPR) on a single Extraction agent — a `complexity_flag` + `template_conf` route between Claude Haiku 4.5 and Sonnet 4. One prompt to maintain, cheaper than two agents, optimization at routing time. Cross-validated extractions against the KG; low-confidence outputs route to Amazon A2I and corrections flow back through SQS → EventBridge → OpenSearch as new few-shot examples.

Stack

  • Python
  • PyTorch
  • ONNX Runtime
  • AWS Step Functions
  • AWS Lambda
  • Amazon Bedrock — Claude Sonnet 4 / Haiku 4.5
  • Bedrock Guardrails
  • MCP (RAG · DB · KG)
  • OpenSearch Serverless
  • RDS PostgreSQL
  • Amazon A2I
  • SQS · EventBridge · S3 (SSE-KMS)
  • API Gateway
  • SageMaker
  • Azure DevOps
  • DataDog

Scale

  • 3-agent serverless pipeline
  • 4-layer Gatekeeper (L0–L3)
  • DocAligner ONNX + CLAHE
  • EfficientNet-B0 (~20 MB · <250 ms on Lambda CPU)
  • IPR — Haiku ↔ Sonnet routing on a single agent
  • OpenSearch RAG few-shot · grows via A2I corrections
  • PII / prompt-injection guardrailed
02

Gannon University

Graduate Assistant · Oct 2024 — Dec 2025

A RAG advising chatbot the campus actually uses.

  • Built a RAG-based Academic Advising Chatbot from scratch — local LLMs via vLLM, vector DB retrieval, Redis caching, FastAPI backend deployed on GCP Cloud Run.
  • Designed the LangChain agent workflows that route automated student queries; secured the endpoint with Cloudflare tunnels and a JWT proxy at sub-100 ms latency under load.
  • Maintained 99.9% uptime and 80% GPU utilization on the local inference rig — local LLM serving that the lab and students actually rely on.

Stack

  • vLLM
  • FastAPI
  • GCP Cloud Run
  • Redis
  • Vector DB
  • Cloudflare Tunnels
  • JWT
  • LangChain

Scale

  • Sub-100 ms latency
  • 99.9% uptime
  • 80% GPU utilization
  • OpenAI-compatible local LLM
03

Gannon University

Student Research Assistant · Jan 2024 — Oct 2024

CNNs in a metal 3D printer.

  • Applied CNNs to automate powder-bed leveling in Direct Metal Laser Sintering (DMLS) — improving accuracy and reducing material waste in a live manufacturing feedback loop.
  • Co-authored and presented a peer-reviewed paper at NAMRC 53 / MSEC 2025 in Greenville, SC, and a poster at the SAMPE Baltimore-Washington 30th URS at the University of Maryland.
  • Same problem I optimize in production: latency budget, reliability under drift, what to escalate when the model is uncertain.

Stack

  • PyTorch
  • CNN
  • OpenCV
  • DMLS
  • Additive Manufacturing

Scale

  • Peer-reviewed
  • NAMRC 53 / MSEC 2025
  • Poster — University of Maryland
04

BitsKraft (DigiConnect)

AI/ML Engineer · Nov 2020 — Nov 2023

Three production AI systems for a Nepali fintech, each one different enough I had to learn something new.

  • Identity verification — liveness detection + emotion analysis for a video-banking platform. The system that confirms a remote applicant is real before they open a bank account.
  • Signature forgery detection — YOLOv5 model for bank cheque verification at 90% accuracy with 40% fewer false positives, behind a REST API handling 5,000+ requests/day at 99.9% uptime.
  • QR-based voice notification system for POS devices that lifted payment-processing speed by 30% — Azure IoT Hub, Cognitive Services, and Functions doing the work.

Stack

  • Python
  • PyTorch
  • YOLOv5
  • OpenCV
  • Azure (IoT Hub / Cognitive Services / Functions)
  • REST APIs
  • Docker

Scale

  • 5,000 daily requests
  • 99.9% uptime
  • 95% verification accuracy
  • 90% forgery detection
  • 40% fewer false positives
  • 30% payment-speed lift
04 / Decision

L4 — published, peer-reviewed

Research, cited the way I'd cite it myself.

[1] Dahal, R., Zhou, L., & Ji, X. (2025). Automatic Powder Bed Leveling for Direct Metal Laser Sintering Based on Machine Learning. ASME MSEC 2025, Greenville, SC. Paper No. MSEC2025-152596.
  • ↳ NAMRC 53 / MSEC 2025 — Greenville, SC.
  • ↳ Poster: SAMPE Baltimore-Washington 30th URS, University of Maryland.

Putting computer vision into a manufacturing feedback loop is the same problem I optimize in production: latency budget, reliability under drift, and what to escalate when the model is uncertain.

Certifications

  • Fundamentals of AI Agents Using RAG and LangChain
  • IBM AI Engineering with Python, PyTorch & TensorFlow
  • Deep Learning with PyTorch
  • Generative AI Language Modeling with Transformers
  • Introduction to Deep Learning & Neural Networks with Keras
  • LangChain — Develop LLM-powered applications
  • Machine Learning — Coursera
  • Google Prompting Essentials

// soon

Writing on cost-aware LLM systems, MCP, and what it actually takes to put a Bedrock agent in front of users. Not live yet — paper first, posts after.

Need someone who builds AI systems that behave under pressure?