projects

LLM agents, RAG systems and full-stack apps — each one with its source on GitHub.

FeaturedAI Agents2026

RAG Agent with Citation Grounding

Preventing hallucinations: a RAG agent that verifies every claim against a verbatim source quote.

Requires each generated claim to carry a verbatim quote from a retrieved source - claims with unknown sources, fabricated quotes, unsupported terms, or numbers absent from the quote are removed before the user sees them
Scores confidence from verifiable signals (trust x verification x (retrieval + support) / 2) instead of the model's self-reported confidence - in the demo, one invented claim out of two drops an answer to 0.42 and flags it
Skips the LLM call entirely when knowledge-base coverage is weak, routes to a SearchProvider fallback, and abstains rather than guesses if nothing verifies

Python
LangGraph
langchain-anthropic
Claude
Pydantic v2
TF-IDF
pytest

Links

source ↗

FeaturedAI Agents2026

Multi-Tool Orchestrator Agent

Coordinating complex workflows: plans goals as capability DAGs and runs independent tool steps in parallel.

Plans goals as capability DAGs and fans out one parallel LangGraph Send branch per ready step - in the demo, 8 independent searches ran in 0.26 s wall-clock against 2.02 s of summed step time, a 7.8x speedup (local run)
Routes by capability rather than tool name over a dynamic registry with hot-swap, plugins and a circuit breaker, ranking providers by priority, observed reliability, trust, cost and latency
Applies deny-by-default permission scoping with role grants, argument constraints and per-run scopes, re-checking every call with its real arguments

Python
LangGraph
langchain-anthropic
Claude
Pydantic v2
SQLite
pytest

Links

source ↗

FeaturedAI Agents2026

Human-in-the-Loop Approval Agent

Safe, compliant systems: pauses for human approval when unsure, with a tamper-evident audit log.

Gates every action on six uncertainty signals - low confidence, missing info, inconsistent samples, unparseable output, LLM error, high-value policy - and fails closed
Pauses with a native LangGraph interrupt checkpointed to SQLite and resumed with Command(resume=...), using durable JSON tickets so no side effect happens while a request waits and a pending ticket survives a restart
Validates human input before resume: authorised reviewer, separation of duties, every question answered, bounded amount edits, and per-reviewer approval limits

Python
LangGraph
langchain-anthropic
Claude
Pydantic v2
SQLite
pytest

Links

source ↗

FinSight AI - Personal Finance iOS App

2026

Full-stack & mobile

Local-first iOS personal finance app built with SwiftUI and SwiftData, featuring expense tracking, AI-powered spending insights, and a forward-looking financial simulator. Designed around an Apple Intelligence abstraction layer for on-device AI.

+ highlights

Built a local-first iOS app with SwiftUI and SwiftData - no backend dependency, all data persisted on-device with full CRUD and seed support
Implemented a deterministic financial simulator with daily savings, spending reduction, and savings goal sliders - real-time projection updates with compound growth math
Architected a protocol-backed AIInsightService abstraction layer for Apple Intelligence via FoundationModels - guarded behind compile-time and runtime checks so the app degrades gracefully on unsupported devices

Swift
SwiftUI
SwiftData
Swift Charts
Apple Intelligence
iOS
Xcode

source ↗

Structured Output Agent

2026

AI Agents

Making LLMs reliable, not random: forces output into Pydantic schemas and retries with targeted error feedback.

+ highlights

Parses and validates every model reply against a Pydantic schema - invalid output is never returned, only retried or raised as RetriesExhausted
Feeds the rejected reply and exact field errors back on retry - in the offline demo, a non-JSON reply and a schema-violating reply (2 field errors) are fixed on attempt 3
Validates tool arguments before execution and tool results after it, so bad arguments never reach a tool and bad data never reaches the model unflagged

Python
Pydantic v2
LangGraph
langchain-anthropic
Claude
pytest

source ↗

ReAct Planning Agent

2026

AI Agents

Agents that don't infinite loop: a ReAct agent bounded on every axis that degrades gracefully.

+ highlights

Enforces five independent stop conditions - iteration cap, repeated-call detection, wall-clock deadline, tool-error streak, parse-failure streak - each covered by a test
In the infinite-loop demo the model repeats one call with a cap of 20 iterations; the loop guard warns on call 2 and stops on call 3 (3 of 20 turns used)
Gates final answers through a self-critique step with bounded revisions, and surfaces low-confidence acceptance in warnings, never silently

Python
LangGraph
langchain-anthropic
Claude
Pydantic v2
pytest

source ↗

Memory-Enabled Conversational Agent

2026

AI Agents

Agents that remember users: a short-term buffer plus long-term vector recall that persists across sessions.

+ highlights

Combines a verbatim short-term buffer with long-term memory, where durable facts are extracted by the LLM after each turn, embedded, and recalled in later sessions
Keeps context bounded - in a measured 100-turn conversation the raw transcript is about 2,290 estimated tokens while the buffer never exceeds about 393, with older turns folded into a rolling summary
Ranks recall by similarity plus recency decay, importance and use frequency, with a similarity gate so an important memory never surfaces for an unrelated question

Python
LangGraph
langchain-anthropic
Claude
Pydantic v2
SQLite
pytest

source ↗

Cost-Aware Agent Router

2026

AI Agents

Reducing infra costs: routes each task to the cheapest capable model and caps spend per task.

+ highlights

Cut cost 43.7% in the bundled demo (simulated models, 36 tasks): routed cost $0.0646 vs $0.1148 always-premium, with 14% of tasks escalating
Scores prompt complexity with a zero-cost heuristic (length, reasoning keywords, code, step count) to pick a starting tier from fast, balanced or premium with no LLM call
Runs a confidence-gated cascade where a cheap model answers first and the router exits early at the threshold or escalates one tier at a time

Python
LangGraph
langchain-anthropic
Claude
Pydantic v2
pytest

source ↗

ShieldBase Insurance Assistant

2025

AI Agents

AI-powered insurance chatbot with RAG knowledge base, LangGraph orchestration, and multi-step quotation flow for auto, home, and life insurance.

+ highlights

Built a multi-node LangGraph StateGraph orchestrating intent detection, retrieval, grading, retry, clarification, and response finalization
Implemented RAG pipeline with vector search, chunk grading, query rewriting, and fallback clarification using Groq LLM and OpenRouter embeddings
Designed a stateful multi-step quotation flow (identify → collect → validate → quote → confirm) with mid-quote side-question routing to RAG and resume

Next.js
LangGraph
LangChain
TypeScript
Groq
RAG
AI/ML

source ↗