Obsessed with uncovering what happens inside an LLM as it turns one thought into the next.
|
WatchItLearn — Real-Time Reinforcement Learning Visualizer A from-scratch Tabular Q-Learning system that makes an AI agent's learning process observable in real time. Implements an MDP grid-world environment with precomputed integer transition tables and NumPy-optimized training. Streams live agent movements, Q-value heatmaps, exploration decay, and Bellman updates over an asynchronous WebSocket engine. Uses Breadth-First Search (BFS) as a mathematical optimality baseline, with verified convergence to the shortest path across 20 randomized seeds. Achieves 2,185+ training attempts/sec without heavyweight RL frameworks.
|
|
Peitho — Negotiation Intelligence API Replaces black-box LLM negotiation with fully deterministic pricing logic. Bayesian WTP estimation with real-time posterior updates. Phase-gated concession curves across opening, active, and closing phases. Session-aware buyer archetype memory that persists across the full negotiation lifecycle. LLM stays entirely out of the pricing path — structured JSON outputs, zero hallucination.
|
|
ORCA — Agentic AI Marine Decision-Support Platform Multi-agent reasoning system for marine ecosystem intelligence. Integrates Indian oceanographic and weather data sources to generate actionable insights for fishermen, researchers, and coastal authorities. Combines environmental data analysis, collaborative agent workflows, and natural-language interaction to support informed marine decision-making.
|
|
Prompta — LLM Prompt Diagnostics Engine Multi-label DeBERTa classifier detecting five failure modes — Instruction Conflict, Format Pressure, Context Starvation, Scope Ambiguity, Role Collapse — with per-mode thresholds and threshold-gated routing to expensive LLM reasoning. Intent extraction decouples user goal from prompt text before classification. Targeted, traceable rewrites per failure mode. Treats prompts as systems with failure modes, not strings to rephrase.
|
|
Settings Assistant — Intent-Driven Android Settings Navigation Conversational navigation layer for Android Settings. Users describe what they want in Malayalam, Manglish, or English instead of knowing technical setting names or menu paths. Resolves natural-language intent to the relevant control, navigates through the Settings hierarchy, and asks for clarification when the intent is ambiguous.
|
|
Spot Difference — Visual Change Detection Pipeline SAM v1 (ViT-H, 4B params) for automatic instance segmentation — no manual annotations. Multi-signal similarity fusion: OpenCLIP 512-dim embeddings, IoU, shape descriptors, and color histograms via a weighted score matrix. Hungarian algorithm for globally optimal object-pair assignment — not greedy. Confidence-stratified JSON reports, pipeline-integration ready.
|
|
Pragati — AI DPR Analysis Platform Full pipeline: OCR → PDF parse → TF-IDF extraction → XGBoost classifier with confidence scoring. 91% accuracy on feasibility classification. Governance compliance dashboard with real-time scoring and structured report output.
|
stack:
backend: [Python, FastAPI, MySQL, PostgreSQL, SQLAlchemy, Pydantic, Redis, AsyncIO, httpx]
ai_ml: [PyTorch, Hugging Face Transformers, scikit-learn]
cv: [OpenCV, Tesseract OCR, SAM, OpenCLIP]