IbuildAIthatsurvivesproduction.
AI engineer at India’s largest explosives manufacturer, a group that also builds for defence and aerospace. I design tool-calling pipelines, retrieval workflows and agent orchestration that connect models to real APIs, databases and business tools. In the same role I’ve led its computer vision division, running 350+ cameras across its plants, which is where I learned what production actually means.
- Cameras running my models on-premise
- 0+
- Daily safety violations, before and after
- 0<8,000
- Specialised agents in Roost
- 0
- Vehicles read and logged at the gates daily
- 0+
Eight systems, from agents to cameras.
LLM and agent work first, then the vision systems it grew out of. Open any tile for the problem, the decisions and the result.
More vision systems on plant floors
Before LLMs, I owned computer vision from model to deployment across multiple plant sites: on-premise, often with no internet, where a false alarm costs a supervisor’s trust.
Pilot
Headcount AI
DeepStream with NvDCF re-identification, a 20-camera pilot aimed at 150+ feeds
Production
X-ray defect inspection
Framed as anomaly detection because real defects are too rare to classify
In progress
ProcessGuard
Flags skipped or out-of-order steps in a critical process as they happen
Six agents. One person says yes.
Roost is my autonomous job search system, and it’s live. Here is one job moving through it. Agents hand work to each other through plain code, and the model only runs where judgment is needed.
Running…
- 1Code
Discover
Queries Greenhouse, Lever, Ashby, SmartRecruiters, Recruitee and Workday, plus JSON-LD on career pages. No model involved.
1 job - 2Model
Prospect
Grows a persistent registry of companies worth watching.
- 3Model
Score
Scores the résumé against each job, in parallel. Results are cached in SQLite.
- 4Model
Write
Drafts a cover letter tailored to the job.
- 5Browser
Fill
Playwright, guided by vision, fills the real application form.
- 6You
Approve
Nothing is sent until you say yes, by click or by voice.
Running alongsideVoice control through the Gemini Live API, and a React dashboard showing the live browser.
StackPython, Groq, Gemini, Playwright, React, SQLite.
Six rules, each learned on the job.
Each one comes from something I shipped.
- 01
Codecoordinates.Modelsdecide.
Every LLM call should be paying for a judgment. Routing, retries, fan-out and hand-offs are ordinary code, which makes them cheap, testable and boring in the right way.
Where I did thisRoost’s six agents talk through deterministic code, and its discovery agent makes zero model calls.
- 02
Nothingirreversiblewithoutaperson.
Agents that act in the world need an approval point placed exactly where the action can’t be undone. Everywhere else they should move without asking.
Where I did thisRoost fills the whole application, then waits for a click or a spoken yes.
- 03
Accesscontrollivesbelowtheprompt.
A system prompt is not a permission system. Filter what the model can see by who is asking, before the model sees it.
Where I did thisThe plant telemetry chat filters context by RBAC role ahead of the model call.
- 04
Theinterfaceispartofthesystem.
Retrieval quality doesn’t matter if the people who need the answer can’t ask the question. Build for how they actually work.
Where I did thisMaintenance RAG is voice-first and bilingual because the crew’s hands are full.
- 05
Assumetheproviderwillfail.
Rate limits, outages and slow responses are normal operating conditions. Design for them on day one and cache anything you’ve already paid for.
Where I did thisKey rotation across providers and a persistent result cache in Roost.
- 06
Measurethenumberthebusinessfeels.
An accuracy score is a checkpoint. The result is whatever changes on the floor because the system exists.
Where I did thisPPE detection is reported as violations per day, about 8,000 down to under 300.
Tools I’ve actually shipped with.
In production.
Shipped and maintained for real users
- Tool / function calling
- RAG
- Vector embeddings
- LangChain
- STT / TTS
- VLMs
- FastAPI
- PostgreSQL
- MinIO
- YOLO
- NVIDIA DeepStream
- React
- Next.js
- OIDC SSO + RBAC
Built with.
Used end to end in my own projects
- Multi-agent orchestration
- Groq
- Gemini
- Gemini Live API
- Playwright
- SQLite
- Flask
- Node.js
- MongoDB
- PyTorch
- TensorFlow
Classic ML.
Internship projects, 2024
- Time-series analysis
- Price prediction
- Classification
- ABC customer analysis
- Survival analysis (Cox-PH)
- Churn prediction
- NLP extraction
- Power BI
- Tableau
From plant floors to agents.
2021–25
B.Tech in Artificial Intelligence
GH Raisoni College of Engineering, Nagpur
BioMedical Hackathon finalist (2022). Smart India Hackathon and Kavach cybersecurity hackathon (2023).
Jun – Dec 2024
Data Analysis Intern
Solar Industries India Ltd., India’s largest explosives manufacturer
A run of smaller ML and deep learning projects that built the fundamentals, and one big one: the PPE detection pipeline across 350+ cameras that’s still running today.
- Time-series analysis
- Price prediction
- Classification
- ABC customer analysis
- Churn prediction with Cox-PH, 85%+
- NLP data extraction
- PPE detection
May 2025 – now
AI Engineer
Same organisation, full time
One role, two halves. Led the Computer Vision division in IIoT: model architecture, training and on-premise deployment across plant sites, shipping ANPR, X-ray inspection and the Headcount AI pilot. Now, with the AI/ML team, architecting LLM-integrated multi-agent systems and tool-calling pipelines over internal APIs and data.
- LLM agents
- Tool calling
- RAG
- Voice interfaces
- Computer vision
- On-prem deployment
2026
Built Roost, independently, end to end
Six agents, six ATS integrations, a person approving every submission. Live at roost.anuragworks.in.
Let’sbuildsomethingthatships.
I’m open to AI Engineer, Applied AI Engineer, LLM / Agents Engineer roles. Full-time or contract, remote or relocation. Based in Nagpur, India, IST, UTC+5:30.