What a system keeps, what it forgets, and what it costs to be wrong about either. Predictive memory orchestration and contradiction-aware stores.
Getting the right context in front of the model, and knowing when you haven't. Hybrid dense-plus-lexical retrieval and confidence-based routing.
Systems that recognise their own failure to progress and halt, rather than burning tokens producing the appearance of work.
Hybrid retrieval RAG framework
A retrieval framework that fuses dense and lexical signals with Reciprocal Rank Fusion, benchmarked on legal and compliance documents across five custom metrics including faithfulness and cross-document reasoning.
Predictive memory orchestrator
A predictive memory orchestrator pairing a multi-modal next-context transformer with a PPO-trained cgroup v2 tier policy. Beat the Android App Standby Buckets baseline on both warm rate and latency at a fixed 400MB budget.
Contradiction-aware shared memory server (MCP)
An MCP memory server that treats a changed fact as evolution with preserved history rather than an overwrite — aimed at a benchmark gap where every system evaluated scores in single digits.
Semantic circuit breaker for AI agents
A detector that trips on semantic information stagnation rather than textual repetition, so autonomous agent loops can be stopped when they stop making progress — not merely when they start repeating themselves.
Mechanisms from the work above, rebuilt as things you can drag. Free, no signup — if one of them saves you an afternoon, it did its job.
Jan 2026 — Present
AI Club, VIT Chennai
Feb 2026 — Present
AWS Cloud Club, VIT Chennai
Jun 2026 — Present
Viha Online
Repetition detection is the standard guard against stuck agent loops, and it fails in both directions. Building Plateau meant measuring stagnation in embedding space instead — and calibrating the threshold per run.
Why hybrid retrieval fuses ranks instead of scores, what the k constant in RRF really does, and where the method quietly fails on legal and compliance corpora.
Building a predictive memory orchestrator meant choosing an observation space for the policy. Free-memory counters looked obvious and were wrong. Here's what PSI telemetry gives you instead.
Get in touch
Open to internships, research collaboration, and conversations about memory and retrieval systems.