Amal R S

The Problem

Long-running LLM agents accumulate context indiscriminately — old task states, irrelevant tool outputs, and contradictory facts pollute the active window, degrading both reliability and predictability.

What I'm Building

A governance layer that sits between the agent's task executor and its memory store. It intercepts read/write calls and applies relevance filters, recency scoring, and task-conditioned rules to decide what actually gets retained.

Separating short-term context (current task scratch), episodic memory (summarised past runs), and semantic memory (domain facts) into distinct stores with explicit promotion/eviction policies.

Research Direction

Targeting multi-agent systems where memory bleed between agents is a security concern. Explainability is a first-class requirement — every memory decision should be auditable.

The Problem

Students juggle subjects, deadlines, and study habits with no intelligent layer to help them prioritise. Existing apps are either too generic or demand manual effort that busy students don't have.

How It Works

User inputs (subjects, due dates, self-assessed confidence) are structured into prompt templates. The LLM backend generates a prioritised study plan and surfaces contextual tips based on topic difficulty and time remaining.

The architecture cleanly separates AI logic from application logic — the LLM is a replaceable backend, not a tightly coupled dependency.

What I Learned

Prompt engineering for structured output is harder than it looks — small wording changes cause large behavioural shifts. I built a lightweight eval harness to catch regressions when prompts changed.

System Architecture

Three tightly coupled subsystems: the flight controller (C++ firmware running on embedded hardware), a Python coordination layer handling sensor fusion and state estimation, and an Android app (Kotlin) consuming the RTSP stream.

Key Challenges

Latency in the video pipeline was the hardest problem — RTSP over WiFi introduced variable frame delay that made gyroscopic VR control feel disconnected. We reduced end-to-end latency by tuning the RTSP server buffer settings and offloading decode to hardware on-device.

Sensor drift during hover required a complementary filter combining accelerometer and gyroscope data — a simple but effective solution that avoided the overhead of a full Kalman filter.

What I Learned

Real-time embedded systems tolerate no abstraction leaks. I became deeply familiar with memory layout, interrupt handling, and the cost of crossing hardware/software boundaries — lessons that now inform how I think about AI agent latency.