Prem
Kumar.
AI Engineer building agentic systems at Ancileo and shipping SaaS products that make work simpler.
Who I am
and what I build
At Ancileo I build the agents that process travel insurance claims end-to-end. LangGraph, RAG over policy docs, runs on Azure. 80% of claims go through without a human.
On the side: the Simpler suite (invoices, outreach, disputes, infographics). Scratch-my-own-itch projects that became actual products. I like systems that are boring to operate.
Where I've worked
and what I shipped
- Currently building a conversational insurance agent using OpenAI's Agents SDK and Strands — multi-turn, context-aware, handles the full claims conversation end-to-end. Claude Code does most of the heavy lifting on the engineering side.
- Built an earlier claims agent on Azure using LangGraph and RAG over policy docs — ran without human-in-the-loop for 80% of cases.
- Cut claims processing time by 60% and kept it running at $2/hour by optimising token usage and caching intermediate steps.
- Ran structured evals across Mixtral, Qwen, and Llama using DSPy and LlamaIndex to find what holds up in production, not just benchmarks.
- Built task-based AI agents for enterprise clients — the kind that replace a repeatable human workflow, not just answer questions.
- Shipped the first RAG pipeline at the company: document ingestion, retrieval, summarisation, hooked into an LLM chatbot interface.
- Wired agents into WhatsApp and Gmail. Customer engagement went up 35% — though that number is partly a product decision, not just engineering.
- Set up LLM-Ops on AWS: LangSmith for tracing, Grafana and Prometheus for infra. The goal was catching regressions before users did.
- Ran evals with RAGAS on synthetic data across model versions. Got output quality up 35% by cutting hallucination on retrieval-heavy tasks.
- Built passive liveness detection for eKYC — the model tells a live face from a printed photo or screen replay without asking the user to blink or nod.
- Shipped 2D/3D face detection (3DDFA) as a SaaS service on TensorFlow. Small team, fast iteration.
- Built an internal dashboard for tracking model accuracy across eval datasets — made regressions visible before they hit production.
- Read ML papers and turned the useful bits into production improvements. Most papers don't survive contact with real data.
- Worked on electronics for Dyson's robotics division across Singapore and Malaysia — the hardware side, not the app.
- Designed and tested circuit boards. Slow feedback loops, a lot of patience.
- Wrote Python scripts to automate hardware testing. Saved a meaningful amount of manual QA time.
- Built liveness detection for e-KYC used by private and government clients in Malaysia — the thing that stops someone holding up a photo to a camera.
- Trained facial recognition models on DepthMap, 3DDFA, and FaceNet. A lot of the work was data, not architecture.
- Sourced training data from 30+ datasets across countries. Diversity in training data matters more than most ML engineers admit.
- Got the company ISO 30107-3 certified — first in Southeast Asia at the time. Mostly paperwork, but the underlying model had to actually meet the bar.
- Built a palm seed classifier for Sime Darby Plantation — multi-class CV model to sort seeds by quality at scale.
- Used AlexNet as the backbone with an SVM on top as the final classifier. This was 2017; that was the move.
- Turned the results into visualisations for non-technical stakeholders. Learned early that a model nobody understands doesn't get deployed.
Things I've built
and shipped to real users
Hiregents
Pick an agent. Connect your Telegram bot. Agent is live and running on its own private server in minutes.
Simpler Invoices
Invoicing for freelancers who don't want to think about invoicing. Generate, send, and track — the whole thing takes under a minute.
Simpler Infographics
Drop in a PDF, get a print-ready infographic. Three output styles — pick the one that fits the audience.
Simpler Outreach
Cold email that doesn't read like cold email. Personalised at scale using Claude Haiku — fast enough to not be annoying.
Simpler Disputes
Chargeback dispute letters, generated. The process is tedious by design — this makes it take two minutes instead of two hours.
Other noteworthy projects
Catering Chatbot
LangChainOpenAIReactOrder-taking chatbot for Caterspot. Handles the full flow — menu, quantities, delivery — so staff don't have to repeat the same questions.
Get in touch
let's build something
I'm open to consulting engagements and product partnerships. Whether you want to talk about AI systems, explore a collaboration, or just say hello — reach out.