AI Engineer — Full-time roles & freelance builds · accepting projects
AI agents, RAG systems, guardrails, ML models, and full-stack products — taken from first scope to deployed, tested, and monitored.
Available for freelance and contract work alongside full-time conversations. Every service below is backed by shipped projects — scroll down for the proof.
Custom assistants for your product or operations — multi-agent workflows, tool-calling, voice agents, web bots. Built with state machines and fallbacks so they behave in production, not just in demos.
Chat with your documents, grounded answers over your data, semantic search inside your app. Six RAG systems built across four vector stores — plus a harness that measures which retrieval setup actually wins on your data.
Prompt-injection defence for your LLM features, security scanning of your web app, and EU AI Act readiness. I have built the guard layer twice and validated the scanner against live production domains.
Prediction, matching, forecasting, recommendation — plus the dashboards to see it. From a TF-IDF product matcher running in production to full MLOps loops with drift detection and automated retraining.
MVPs, internal tools, and complete platforms — backend, frontend, database, deployment. Comfortable across Python, Go, and TypeScript, with real CI/CD and tests, not just a working demo.
Turn documents, websites, and manual workflows into structured data and automated pipelines — OCR on invoices and certificates, scraping with headless-browser fallbacks, scheduled jobs and alerting.
Typical engagements: a scoped build (2–8 weeks), an audit, or ongoing retainer work. Tell me what you need →
The flagships — proof of everything listed under services. 13 shown of 58 built across open source, research, coursework and client work; the public record lives on GitHub ↗.
Open-source AI governance, risk & compliance platform. Classifies AI systems against EU AI Act risk tiers with statutory citations, generates required documentation, defends live LLM endpoints with a five-layer prompt-injection guard (regex + fine-tuned DeBERTa), and answers regulatory questions with groundedness-scored RAG.
A 20-check DAST engine for web security scanning, validated against 53 live scans of Indian government production domains — surfacing 77 critical and 236 high-severity findings.
Seven LangGraph agents cooperating over a Redis blackboard to triage security events — MITRE ATT&CK technique mapping, CVSS v3.1-to-v2 fallback parsing, per-agent exception isolation so parallel dispatch can't cascade.
Standalone five-layer defence pipeline: Unicode NFKC normalisation with 49-entry homoglyph canonicalisation → severity-graded regex families → fine-tuned DeBERTa-v3-small intent classifier → a named-rule decision ladder (0.4 regex / 0.6 intent) → recovery-oriented sanitiser. Mapped to OWASP LLM01/LLM02/LLM07.
Risk-tier classification as an executable Article 5 → Annex III → Article 52 precedence cascade — a 16-signal pure-function rules engine chosen over an LLM, because a regulator must be able to replay every decision. Cross-walked to NIST AI RMF subcategories.
Natural-language analytics over a production Databricks warehouse — with a deterministic SQL firewall between GPT-4 and the data: a statement-prefix whitelist and nine-verb deny-list the model cannot influence.
A complete champion/challenger model lifecycle inside Snowflake: warehouse-native features that eliminate training-serving skew, a three-model bake-off, daily scoring, weekly retraining gated on live production error, KS drift detection, and atomic model cutover.
A 10-page IEEE-format audit of income-prediction models across 29 Gender × Race × Age subgroups — nine of which received zero positive predictions at DIR 0.00 against the legal 0.80 four-fifths threshold. Independently verified and corrected a third-party fairness-metrics library along the way.
A PPO agent for the board game Risk with a variable action space of hundreds of legal moves — GAE, clipped surrogate objective, orthogonal initialisation and linear LR annealing written by hand, trained to 1,000,000 timesteps. Evaluated in a seat-swapped round-robin tournament that controls for first-player advantage.
A controlled four-arm experiment built so retrieval architecture is the only variable — with a real 80-row results CSV including negative results, reported honestly. One of six distinct RAG systems in the portfolio spanning four vector stores.
Multilingual RAG customer-support agent (final-year major project) — a substantial SPA with no build step, a three-tier knowledge-source cascade with explicit fallbacks, and 3,000+ lines of architecture documentation treated as a first-class deliverable.
Power BI dashboard analysing investment behaviour across demographics, income bands and goals — interactive drill-downs, DAX measures and segment comparisons, built to be read by non-technical stakeholders.
Advanced SQL analysis over financial transaction data — multi-table joins, subqueries, window functions and query optimisation, with schema normalisation for reporting workloads.
Production AI for sustainability intelligence — shipped inside a live multi-service platform used by real customers.
Core engineer on an enterprise AI platform for carbon footprinting, Life Cycle Assessment and supply-chain intelligence — 16 containerised FastAPI microservices backed by PostgreSQL, Redis, MongoDB and AWS, shipped to four environments through a per-service change-detection CI/CD pipeline.
A ~26,000-line multi-agent subsystem that assembles Digital Product Passports conversationally — an explicit state-machine coordinator and a declarative save-operation registry where every backend write flows through one persistence choke point: a save is reported only if the backend actually accepted it, unchanged values are never re-sent, and every failure records the backend's own error code.
A four-stage URL-import pipeline for the passport agent: SSRF-guarded, size-capped fetch with headless-render fallback → static parse of structured markup (no LLM) → one JSON-mode LLM pass over the cleaned signals → resolution against internal catalogues, where anything unmatched is flagged for support — never silently guessed.
Re-architected a production chatbot to hybrid BM25 + vector retrieval with cross-encoder re-ranking and Redis caching; then built its analytics product — persistent conversation storage, a dashboard that federates four isolated environments (direct DB reads where reachable, service-to-service HTTP where not), CSV export, and automated alert emails for unanswered questions.
The platform's decisioning pattern: an LLM proposes a selection from database candidates, and a deterministic guard layer disposes — enforcing domain invariants like flag gating, dedup and mandatory steps — with an AI circuit breaker that falls back to pure rules when failure rates spike. Deterministic spec-sheet parsers resolve locations against live geography tables and flag what they can't match.
A family of matchers linking messy external data to internal catalogues: hierarchical SKU matching (exact → external ID → TF-IDF cosine) streaming batch results over Server-Sent Events with config-driven scoring, PO line-item matching, n-gram material matching, and a four-tier process cascade (exact → fuzzy → AI-assisted → flagged fallback). Plus queue-driven document extraction from purchase orders, BOMs and compliance certificates with field-level confidence scoring that routes low-confidence documents to human review.
Browser voice agents on the OpenAI Realtime API using the ephemeral-key WebRTC pattern — audio never proxies through the backend — with RAG tool-calling over a vector knowledge base and helpdesk ticketing, including an OAuth2 vendor migration. Plus a narrated-video pipeline (speech-to-text → TTS → avatar render → ML background segmentation → FFmpeg composition) with an engine pin that prevents a 4× render-cost regression.
problemA sustainability platform's ops team hand-matched thousands of retailer products to internal footprint data.
builtA hierarchical matching engine — exact → ID → text similarity — streaming batch results live, every threshold config-driven.
outcomeMatching runs unattended; ops re-tune scoring themselves, no engineering release needed.
problemA production chatbot answered most questions — nobody knew which ones it missed, or for whom.
builtHybrid retrieval plus a federated analytics dashboard across four isolated environments, with alert emails on unanswered questions.
outcomeCoverage gaps became a metric; missed questions reach the team's inbox the moment they happen.
problem"How exposed are we, actually?" — the question every public web app owner avoids.
builtA 20-check DAST scanning engine, validated against 53 live production domains.
outcome77 critical and 236 high-severity findings surfaced — before an attacker found them.
End-to-end data science fellowship work: cleaning, feature engineering, and model development on real datasets alongside senior data scientists — culminating in a 14-week portfolio of 100+ algorithm implementations, from classical ML through deep learning and NLP.
Grouped by where it shows up in the work — every item below is backed by a shipped project.
Python
TypeScript
JavaScript
Go
C++17
C
SQSQL
Dart
Kotlin
Swift
Solidity
Bash
PowerShell
HTML5
CSS3
OWASP LLM Top 10
EU AI Act
NINIST AI RMF
MAMITRE ATT&CK
GDGDPR
DADAST
FAFairness Auditing
AIAIF360
SHSHAP
LILIME
JWT
OAuth 2.0
AEAES-256-GCM
PApip-audit
Claude Code
Claude API
OpenAI GPT-4o
Gemini
GQGroq
Llama 3
Ollama
DeepSeek-R1
Qwen
Mistral
Hugging Face
BRAWS Bedrock
TATogether AI
vLLM
LM Studio
Whisper
ElevenLabs
LangChain
LangGraph
CrewAI
MCP
A2A2A Protocol
RARAG Pipelines
MUMulti-Agent Systems
TOTool Calling
RVRealtime Voice
Qdrant
ANAnnoy
Elasticsearch
STSentence-Transformers
OpenAI Embeddings
TKtiktoken
PMPyMuPDF
TETesseract OCR
PyTorch
TensorFlow
Keras
scikit-learn
XGXGBoost
Transformers
ONNX Runtime
UNUnsloth / QLoRA
SBStable-Baselines3
Gymnasium
OpenCV
YOLOv8
MediaPipe
TensorFlow Lite
SPspaCy
NLNLTK
NumPy
pandas
SciPy
NetworkX
Matplotlib
SESeaborn
Plotly
Jupyter
Google Colab
Streamlit
MLflow
TBTensorBoard
Roboflow
PostgreSQL
Neon
MySQL
SQLite
MongoDB
Redis
Snowflake
Databricks
DuckDB
Parquet
SASQLAlchemy
ALAlembic
Prisma
AWS S3
Cloudflare R2
Tableau
Power BI
FastAPI
Django
Flask
Pydantic
Node.js
Fastify
Next.js
React
Svelte / SvelteKit
Vite
Tailwind CSS
shadcn/ui
Radix UI
ZUZustand
TanStack Query
React Router
React Hook Form
Zod
Three.js
D3.js
Chart.js
RERecharts
Swagger / OpenAPI
WebSockets
Stripe
Firebase
Twilio
Selenium
Playwright
React Native
Flutter
Android
iOS
ESP32 / ESP-IDF
FRFreeRTOS
raylib
Chrome Extensions
Ethereum
Polygon
Web3.py
IPFS
MMMetaMask
AWS
AWS Lambda
AWS EC2
Google Cloud
Docker
Kubernetes
GitHub Actions
Git
GitHub
Linux
Ubuntu
Nginx
Vercel
Netlify
Render
Heroku
Cloudflare
Let's Encrypt
Prometheus
Grafana
OpenTelemetry
Sentry
PyPI
npm
pnpm
uv
pytest
Vitest
Jest
Postman
ESLint
Prettier
Black
MYmypy
pre-commit
VS Code
Specialised in Data Science: machine learning, deep learning, NLP, DBMS, statistics, and data structures & algorithms.
Physics, Chemistry, Mathematics and Computer Science.
Lead the technical community: event planning, workshops, hackathons. Previously Technology Trainer (2023–24), running sessions on programming and data structures.
First place for an AI/ML solution built under competition constraints.
Two delivered talks on AI literacy and the GenAI threat landscape, including a live synthetic-media deception demo.
Community development and social-welfare initiatives supporting underprivileged communities; regular anchoring and public speaking at college and community events.
Free, ~30 min. You describe the problem; I tell you honestly if and how it's buildable.
Written scope, timeline and price before any commitment. No open-ended billing.
Working software every week — not a big reveal at the end. Course-correct early.
You own everything: repository, tests, deployment docs, and a walkthrough.
remote-first · IST (UTC+5:30) — overlaps EU & US mornings · replies within 24h
I take on freelance and contract projects — AI agents, RAG systems, automation, ML models, security audits, and full-stack builds — alongside conversations about full-time AI / ML roles. Send a couple of lines about what you need and I'll reply with a scope, a timeline, and an honest feasibility take.