Llm

Dialectic is an adversarial reasoning application built to improve decision quality on complex questions by forcing two AI agents to debate opposing sides instead of generating a single one-shot answer. GitHub: github.com/singhaditya8499/Dialectic What We Built We built a complete end-to-end debate platform with: Two-agent structured debate loop (for vs against, or A vs B) Provider-agnostic model layer supporting OpenAI, Anthropic, and Ollama Live streamed debate transcript over SSE for real-time UX Evidence-anchored turn schema (facts, figure/date, source, reliability) Structured summary pipeline that preserves major claims and evidence ledger Local debate archive with reload support from the UI Browser-compatible mode for GitHub Pages deployments Product and Technical Design Core architecture src/server.js: HTTP API, SSE stream endpoint, static asset serving, persistence APIs src/debateEngine.js: debate orchestration, turn normalization/validation, summary generation src/providers.js: common completion abstraction across providers public/*: setup form, transcript renderer, summary view, saved debate browser Debate lifecycle User submits question, side models, and round budget. Engine alternates turns between both sides. Each turn is validated for stance, rebuttal quality, novelty, and evidence quality. Streaming events update UI in real time (thinking, turn, summary, complete). Final structured summary and debate artifact can be saved locally. Quality and Alignment Guardrails Dialectic focuses on controlled generation rather than raw free-form output: ...

CardWise AI is a personal-finance assistant focused on helping users evaluate credit-card options, align card selection with financial goals, and understand spending behavior through AI-generated summaries. Focus Areas Personalized card recommendations based on user goals and constraints. Goal-aware planning support for safer and more structured decision-making. Natural-language insights to explain spending patterns in simple terms.

Overview DocCompass is a local-first document discovery system built to find personal files by meaning, not only by filename. It is designed for natural requests like: “my marksheets” “visa related documents” “intro to psychology notes” The project searches local directories, extracts text from multiple file formats, builds a reusable index, creates a structured query plan, and ranks results by relevance with transparent scoring. Repository: github.com/singhaditya8499/DocCompass Problem Statement Traditional local file search is often fragile because it depends heavily on exact filenames and folder memory. In real usage, users usually remember intent and partial context, not exact paths. DocCompass addresses this by combining: ...

This post outlines prompt and retrieval strategies, evaluation loops, and safety checks that improved enterprise chatbot answer quality from baseline to production-ready levels.

An AI-driven recommendation prototype combining LlamaIndex pipelines, multi-model analysis, and sentiment extraction from financial news streams. Repository/demo assets can be embedded here as the project evolves.