Freelance
LLM engineer for production AI: document automation, data assistants, fine-tuning, and evaluation.
I build AI systems that run in production, and I measure whether they work.
Currently contracting with Synaptics on enterprise AI automation.
- Location
- Remote · Bengaluru, India
- Hours
- 4 pm – 1 am IST
- US time
- 6:30 am – 3:30 pm ET · 3:30 am – 12:30 pm PT
- Availability
- 20–25 hrs/week
US times move one hour earlier from November to mid-March. About 6 hours of overlap with US Eastern business hours.
Track record
Live in Korea
AP Invoice Automation; rollout underway across 8 global regions
Week → days
Week-long manual CAD process replaced by a few-day automated flow
6 + 5
Public models and datasets on Hugging Face
~49,000
Generations evaluated, with human annotation and an LLM judge
Services
AI workflow & document automation
Invoices, forms, and technical files turned into structured data and loaded into your ERP or database, on a schedule.
- Extraction pipeline
- ERP or database integration
- Scheduling and monitoring
See example: AP Invoice AutomationCAD extraction pipeline
LLM apps: data assistants & agents
Ask your business data questions in plain language. Read-only access, role-based permissions, and answers your team can check.
- Chat interface
- SQL or tool-using agent
- Auth and access control
See example: Procurement chatbotConversational Data Analyst
LLM fine-tuning & compression
Adapt open models to your domain with LoRA or QLoRA, and quantize them to run on smaller, cheaper hardware.
- Fine-tuned adapters
- Quantized model
- Benchmark report
See example: FedQuant paperPublic models on Hugging Face
LLM evaluation & training data
Find out whether your model actually reasons correctly, not just whether it lands on the right answer. English and Hindi.
- Evaluation suite
- Annotation guidelines
- LLM-judge setup
- Significance tests
See example: Step-level evaluationCross-lingual reasoning
AI prototypes & MVPs
From idea to a working web app your users can try: FastAPI backend, React or Next.js front end, deployed.
- Working prototype
- Deployed demo
- Path to production
See example: ModelBench, liveAP Invoice POC (Best AI Work)AI Resume Analyzer, live
AI architecture review
A second opinion before you build or scale: architecture, data flow, access control, and running cost, with written recommendations.
- Architecture review
- Risks and fixes
- Written report
See example: AP Invoice enterprise stackModelBench (model selection)
Stack
- Models
- PyTorch, Hugging Face (Transformers, PEFT), LoRA, QLoRA, NF4 / AWQ / GPTQ quantization
- LLM apps
- NL→SQL agents, multi-agent systems (Autogen)
- Document AI
- Azure Document Intelligence, Azure Content Understanding
- Backend
- Python, FastAPI, Redis, Docker, Next.js / React, Oracle SQL
- Cloud & enterprise
- Azure, OCI Functions, Oracle Integration Cloud, Oracle EBS, Alteryx Server
- Evaluation
- Human annotation, LLM-as-judge, McNemar, Wilcoxon, Cohen's d
What you get
- Working code in your repository
- Setup and handover documentation
- A walkthrough call for your team
- Your data stays in your own cloud accounts
Ways to work together
Part-time contract
Ongoing work at 20–25 hours a week, inside your team's tools and rhythm.
Fixed-scope build
One system with an agreed scope, milestones, and delivery date.
Evaluation sprint
Test a model or LLM app, then a report with numbers and concrete fixes.
How I work
We talk about the problem
A short call or email about what you want to change and what data you have.
I write down the scope
What I will build, what I need from you, and how we will know it works.
I build in small steps
You see working results early and can change direction before much time is spent.
You get a clean handover
Code, setup notes, and a walkthrough, so your team can run it without me.
Questions
- Which time zone do you work in?
- India (IST), 4 pm to 1 am. That covers US Eastern mornings and early afternoons.
- Will you sign an NDA? What happens to our data?
- Yes. I work inside your accounts and keep no copies of your data.
- Who owns the code?
- You do, from the first commit.
Have a project in mind?
The button opens an email with three short questions. Answer what you can; a few lines is enough.