The flagships — proof of everything listed under services. Hiring? This is the depth. Freelancing? This is what shipped looks like. Full record 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.
~34,000 LOC · 356 tests · 288 commits · 6 Alembic migrations · K8s HPA manifests · SDK + MCP server + Chrome extension
securityeu ai actFastAPIPyTorchReactFAISS
A publishing platform in Go + Next.js with a hand-written MCP server implementing JSON-RPC 2.0 directly from specification. Hybrid relevance ranking fuses weighted full-text search with trigram similarity and log-damped engagement; PWA push, archive-bomb defence, single-flight token refresh.
~27,550 LOC Go + TypeScript · hand-rolled MCP server · weighted tsvector + trigram hybrid search
platformGoNext.jsPostgreSQLMCP
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.
144 unit tests · 53 live scans · 77 CRITICAL / 236 HIGH findings
offensivePythonDASTAppSec
Cyber Defense SOC Platform
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.
~5,900 LOC · 7 agents · regex pre-filters · temperature 0.1 · bounded fan-out
securityagenticLangGraphRedis
LLM Prompt-Injection Guard
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.
custom Dataset/DataLoader/warmup schedule · max-severity scoring prevents false-positive stacking
securityPyTorchTransformersOWASP
EU AI Act Compliance Engine
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.
deterministic by design · statutory citations on every factor · prohibitions short-circuit at confidence 0.99
governancePythonNIST AI RMF
AI Data Analysis — Text-to-SQL
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.
5,137 lines · non-LLM validation gate · query-scoped tenancy in the WHERE clause
securityGPT-4DatabricksSQL
Snowflake Healthcare MLOps
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.
6,474 lines · 2% improvement threshold so models don't churn on noise · deprecate-then-insert cutover
mlopsSnowflakePythonDrift Detection
Intersectional Fairness Audit
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.
IEEE paper · EEOC four-fifths rule applied · AIF360 dtype failure independently debugged
researchFairnessResponsible AI
Risk Game Bots — PPO from scratch
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.
6,087 LOC · no RL framework · 1M timesteps · controlled evaluation
rlPyTorchPPOPython
Four-Strategy RAG Harness
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.
measurement over assertion · honest negative results · reproducible seeded splits
researchRAGEvaluation
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.
multilingual retrieval · graceful degradation cascade · docs-as-deliverable
agenticRAGFastAPIJS
Investment Survey Dashboard
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.
DAX measures · demographic segmentation · interactive drill-down
analyticsPower BIDAXData Viz
Financial Data SQL Analysis
Advanced SQL analysis over financial transaction data — multi-table joins, subqueries, window functions and query optimisation, with schema normalisation for reporting workloads.
window functions · query optimisation · normalised reporting schema
analyticsSQLWindow Functions
