▸ AI Agents · Document AI · LLM Systems
Building production AI that holds up in the real world.
I take AI from prototype to production, working alongside the teams who rely on it, and build for what happens long after launch.
- AI Engineer
- AI Consultant
- Agentic AI Builder
- Creator
Building AI that powers public services at state scale
- 4+ yrsbuilding production AI
- 19AI systems built
- 10clients served worldwide
- 30K+engineers follow my work
▸ About
Built for production, not just experiments.
I build AI that ships. Over the past 4+ years I have delivered systems that run in healthcare, transport, insurance, enterprise and public services, usually working alongside the teams who use them, turning models into products people genuinely rely on.
Today my work centres on AI agents, search and answering over company documents, and the services behind them that keep everything running once it is live.
What I bring
- Systems handling 812K records in live government use
- Security tooling that proves a flaw is real before reporting it
- 40% productivity lift for an enterprise data-migration team
- Models that lifted insurance renewals past 60% and cut baggage mishandling by 20%
How I work with clients
- AI/ML and GenAI consulting, architecture to production
- Advisory for teams shipping their first AI systems
- Brand collaborations with a 30K+ engineering audience
▸ Work with me
Ways we can work together.
Currently open to select freelance work, advisory, and brand collaborations.
▸ Featured work
Systems already running in production.
Each project below solved a real problem for real users, from the first design through to running live. I bring the same care to client work.
MedAuth AI
Prior-authorization decision support that helps health insurers decide whether to approve a treatment before it happens. RAG retrieval is pinned to the policy version in force on the date of treatment, an AI model checks each rule on its own, and fixed code, not the model, produces the recommendation the human reviewer sees.
- Every quoted line is verified against its source document
- Role-scoped permissions; the database refuses edits to audit records
- 1,400+ tests, plus a fault-injection run that caught all 47 planted safety bugs
- Python
- FastAPI
- PostgreSQL
- pgvector
- RAG
- LLM
LLM Firewall
A security gateway that sits in front of any OpenAI-compatible LLM API and inspects every request and response. It blocks prompt-injection and jailbreak attempts and redacts PII, refusing flagged traffic in about 9 ms before it ever reaches the model.
- Policy engine with four enforcement actions and per-caller service keys
- HTTPS enforced behind nginx, with 16 Prometheus alert rules
- 1,800+ tests, including adversarial ones aimed at the gateway itself
- Python
- FastAPI
- PostgreSQL
- Docker
- nginx
- Prometheus
Pentest Agent
A security testing platform that probes a live website the way a real attacker would, using AI agents that each attack from a different angle. Unlike the standard tools, it only reports a problem after re-running the attack to prove it is real. On a seven-application industry benchmark it found 56 confirmed issues where OWASP ZAP found none.
- Six stages: map the site, attack it, verify each finding, review, then report
- Nothing is reported until the attack has been re-run against the live site
- 12 reports, mapped automatically to 6 compliance standards
- Python
- FastAPI
- Multi-agent
- Next.js 16
- Playwright
- 13 LLM providers
Release Agent
Turns a code push into a full deployment with no human in the loop. It watches for new code, then builds, ships, and health-checks the release on its own, and when something breaks it works out why and rolls back safely. It also scans each change for security problems before anything goes out.
- Connects to both Azure DevOps and GitLab
- Runs across 118 backend services, taking releases from weekly to several a day
- Works out why a release failed and rolls it back automatically
- Python
- Docker
- Strix
- Azure DevOps
- GitLab
- LLM
Response Quality AI
Reads 812K officer replies to citizen complaints for a state government and judges whether each one actually resolves the complaint. A translation model running in-house handles the 438K written in the regional language before an AI model grades the response.
- Graded 283K cases and found roughly 70% left the complaint unresolved
- Cut model calls by 56% by caching text that repeats across replies
- Built to run over a live public grievance system, not a sample set
- Python
- NLLB-200
- LLM-as-judge
- Caching
Document Intelligence
Turns scanned answer documents into clean, structured records in two languages. Work that 40 teams once retyped by hand is now checked and approved in two steps, cutting turnaround from days to hours.
- Reading and extraction models run in-house, not through an outside API
- Two-step review keeps a human in control of what gets published
- Shipped with its live deployment and monitoring
- Python
- OCR
- LLMs
- Prometheus
- Grafana
Report Agent
Writes Detailed Project Reports through a guided conversation. It pulls from large company document stores and backs every section with the source it came from, producing consistent reports and cutting drafting time considerably.
- Builds the report section by section instead of in one pass
- Each section carries the source document it was drawn from
- One of a set of in-house tools that cut report and lookup time by 35%
- Python
- vLLM
- RAG
- Vector DB
- LLMs
Ops Intelligence
The analytics platform behind a statewide operations dashboard and the systems around it for a state government, including its public transport arm. It turns large amounts of day-to-day data into clear answers, replacing manual reporting and speeding up decisions.
- Feeds a statewide operations dashboard and the systems around it
- Replaced manual reporting across multiple departments
- Processes large operational datasets to surface what needs attention
- Python
- ML
- NLP
- PySpark
- BI
Have a problem worth solving?
Tell me what you are building, and I will show you how I would take it to production.
▸ Testimonials
Recommendations from people who know my work.
I had the pleasure of managing Piku for a significant duration, and his performance in AI, ML, NLP, and Python was outstanding. He consistently exceeded expectations, delivering high-quality work within tight deadlines. His strong analytical skills and attention to detail were very much valuable to our team.
I interacted with Piku during his time at Philips. He was a sharp and well-rounded engineer. Although I did not work with him directly, his project supervisor spoke highly of how knowledgeable he is and how consistently he demonstrated Philips values.
▸ Experience
A track record of building AI that runs.
Mar 2024 – Present
CurrentAI Engineer, Centre for Smart Governance (CSG)
GovTech · Public Sector
Build and run AI systems for Karnataka government departments, embedded with the people who depend on them daily. The work spans grievance analysis over 812K officer replies, bilingual document processing for 40 teams, AI-driven security testing, in-house assistants for reports and data, and an autonomous release pipeline covering 118 backend services.
- Python
- vLLM
- NLLB-200
- RAG
- LLMs
- CI/CD
Aug 2023 – Mar 2024
AI/ML Consultant, Independent · via Fiverr
Freelance · Global clients
Delivered AI and data projects for clients around the world as an independent consultant, from document search and AI assistants to machine-learning models, cloud data pipelines, and reporting dashboards.
- Python
- RAG
- LLMs
- ML
- Cloud
Feb 2023 – Jul 2023
Data Developer, Philips
Healthcare · Enterprise
Built a cross-platform desktop tool in Python, MySQL and PyQt5 that let the Data Migration team estimate effort on new projects, lifting their productivity by 40%. Standardising onboarding and automating the validation checks cut errors by 60% and raised efficiency by 30%.
- Python
- PyQt5
- MySQL
- Data migration
Apr 2021 – Sep 2021
Data Scientist, Innodatatics
Insurance · Aviation
Built a health-insurance renewal-prediction app that lifted renewal rates past 60%, plus aviation analytics dashboards that cut baggage mishandling by 20% and luggage losses by 40%.
- Python
- scikit-learn
- ML
- Analytics
▸ Core expertise
Where I go deepest.
RAG & grounding
Ask questions of your own documents and get an answer that cites the source it came from, built to stay dependable once it is live.
AI agents
Agents that plan, use tools, and work together on multi-step tasks. Designed with a clear reasoning flow, limits that keep them safe, and a defined path for when something fails.
Evaluation & reliability
Knowing whether the output is actually right: grading answers at scale, attacking the system on purpose, and planting faults to check they get caught.
Productionising AI
Serving models on your own infrastructure, the APIs around them, and the monitoring that keeps everything standing once real users arrive.
▸ Presence
An engineer who builds in public.
I teach production AI to the engineers building it, with practical breakdowns and honest takes on the parts most write-ups leave out.
▸ Contact
For consulting, freelance builds, advisory, and brand collaborations.
I read every message myself and reply within 24 hours. Tell me a little about what you have in mind.